A method, device, equipment and medium for incremental learning prediction of time series data

Abstract

The present invention provides a method, device, equipment and medium for incremental learning and prediction of time series data, which relates to the field of time series data learning and prediction. The present invention performs sliding window segmentation and feature alignment on the acquired multi-dimensional time series data, and then converts it into a two-dimensional spectrum graph through Mel spectrum analysis; then constructs a self-attention incremental learning framework based on pre-training, and combines the elastic weight solidification module to fix the key parameters of historical tasks; and uses an improved experience replay mechanism to screen out representative samples and complex samples, and save them to the replay buffer; at the same time, integrates the dynamic distillation loss function for training to achieve the fusion of new and old knowledge, and obtain an incremental learning prediction model for subsequent prediction tasks. The present invention solves the problem of catastrophic forgetting in dynamic scenarios of massive industrial time series data through a multi-strategy collaborative incremental learning mechanism, and significantly improves the accuracy and stability of continuous learning of the model in tasks such as equipment status monitoring and fault warning.

Description

Technical Field

[0001] The present invention relates to the technical field of time series data learning and prediction for industrial equipment, and in particular to a method, apparatus, equipment and medium for incremental learning and prediction of time series data. Background Art

[0002] In modern industrial production, monitoring the operating status of industrial equipment and providing fault warnings are key to ensuring production efficiency and equipment safety. To achieve this goal, it is necessary to collect and process large amounts of time series data. This data is often multi-source and heterogeneously distributed, covering a wide range of dynamic information during equipment operation.

[0003] With the rapid development of artificial intelligence technology, time series classification methods based on deep learning models have achieved remarkable results in industrial equipment monitoring tasks. However, existing deep learning models have shown significant limitations when dealing with the ever-increasing amount of multi-source and heterogeneously distributed time series data, especially in dynamic scenarios where they suffer from a serious problem of catastrophic forgetting. Catastrophic forgetting refers to the gradual loss of knowledge about old tasks as the model learns new ones, resulting in the model's inability to maintain stable performance across multiple tasks. This problem seriously affects the efficiency and reliability of industrial equipment condition monitoring and fault warning based on time series data analysis. In addition, existing technologies often lack effective incremental learning mechanisms when processing massive amounts of time series data, making it difficult to achieve real-time updates and expansions while ensuring model accuracy.

[0004] Therefore, how to design an incremental training method that can effectively solve the problem of catastrophic forgetting and adapt to the characteristics of multi-source heterogeneous distribution data has become a technical problem that needs to be solved urgently.

[0005] In view of this, this application is hereby filed. Summary of the Invention

[0006] The present invention aims to provide a method, apparatus, device and medium for incremental learning and prediction of time series data, achieving technological breakthroughs through time series feature alignment, incremental learning framework construction, improved sample experience replay and dynamic distillation loss function integration, thereby improving the model's continuous learning ability in dynamic scenarios and improving the continuous learning accuracy and stability of industrial equipment models in tasks such as equipment status monitoring and fault warning.

[0007] In order to solve the above technical problems, the present invention is implemented through the following technical solutions:

[0008] A time series data incremental learning and prediction method, comprising:

[0009] S1, obtains multi-dimensional time series data of industrial equipment with multi-source and heterogeneous distribution;

[0010] S2, performing sliding window segmentation and feature alignment on the multidimensional time series data, and converting it into a two-dimensional spectrogram through Mel spectrum analysis;

[0011] S3, inputting the two-dimensional spectrogram into a pre-trained basic model for incremental learning training, and fixing the key parameters of the basic model's historical tasks based on the elastic weight solidification module, and constraining the parameter update of the basic model by adding a regularization term to reduce the deviation from the important parameters of the previous task;

[0012] S4, during the incremental learning training process, calculates the representativeness and uncertainty of the samples in real time, selects representative samples and complex samples, and saves them to the playback buffer for sample playback during the next training;

[0013] S5, combining the current task data with the playback samples in the playback buffer, based on the dynamic distillation loss function, realizes the fusion of new and old knowledge through inter-layer feature distillation to obtain an incremental learning prediction model;

[0014] S6, processing the newly acquired multi-dimensional time series data of the industrial equipment into a two-dimensional spectrum graph and inputting the two-dimensional spectrum graph into the incremental learning prediction model for prediction, thereby obtaining a prediction result of the operating status of the industrial equipment.

[0015] Preferably, the parameters of the Mel spectrum include: jump length, window length, window type and number of Mel filters;

[0016] The S2 is specifically:

[0017] Segmenting the multidimensional time series data according to a preset jump length and window length;

[0018] Perform feature alignment on the segmented data;

[0019] Based on the set window type, the data after feature alignment in each window is Fourier transformed to convert it from the time domain to the frequency domain to obtain a linear spectrum;

[0020] Filtering the linear spectrum through a set number of Mel filters to obtain a Mel spectrum;

[0021] A logarithmic operation is performed on the Mel spectrum to compress the dynamic range of the spectrum, thereby obtaining a logarithmic Mel spectrum, which is visualized as a two-dimensional spectrum graph.

[0022] Preferably, the method further comprises: filtering out rows with null values ​​in the two-dimensional spectrum graph before the two-dimensional spectrum graph is input into the basic model;

[0023] The basic model is a deep learning model with hierarchical time-frequency window self-attention, including a partitioning module, a position embedding module, a hierarchical multi-head self-attention structure module and a window semantic module;

[0024] The division module is used to divide the two-dimensional spectrum into equally spaced windows, and each window is divided into a number of time-frequency blocks;

[0025] The position embedding module is used to add position embedding information to each time-frequency block;

[0026] The hierarchical multi-head self-attention structure module is used to perform self-attention calculation on the time-frequency blocks after position embedding in each layer of the multi-head self-attention structure, and perform time-frequency dimension compression and channel dimension expansion while extracting the relationship between blocks;

[0027] The fully connected layer window semantic module is used to map the data after the channel dimension is expanded to the category semantic dimension through single-layer convolution and full connection.

[0028] Preferably, the elastic weight solidification module is used to fix the historical task key parameters of the basic model to prevent catastrophic forgetting in incremental learning, specifically:

[0029] After each task training is completed, the importance score of the parameters in the basic model to the task is calculated through the Fisher information matrix. The calculation formula is:

[0030] ;

[0031] in, Score the importance of the t-th task; is the learning training data for the tth task; is the i-th parameter of the basic model; For the input sample data of the basic model Output category label of For parameters The basic model of is the true category label; Indicates derivative.

[0032] Preferably, when calculating the representativeness of a sample, it is reflected by the distance between the sample and the class prototype of the task category. The class prototype distance is calculated as follows:

[0033] First, for each category in each task, the feature mean of the data in the task is calculated as the class prototype of the category. The formula is:

[0034] ;

[0035] in, For category The characteristic mean of For category The sample set of is the sample data, is a feature extractor based on the basic model;

[0036] Calculate the cosine similarity of all samples to the class prototype. The formula is:

[0037] ;

[0038] in, for arrive The cosine similarity of Represents sample data The characteristic mean of .

[0039] Preferably, when calculating the uncertainty of samples in real time and filtering out complex samples:

[0040] First, different data perturbations are applied to the input training data, and the model prediction results are obtained based on the perturbed data;

[0041] The uncertainty of sample data is calculated by the stability of the prediction results, and complex samples are screened out; uncertainty The expression is:

[0042] ;

[0043] in, is the sample data; Task collection; For the current task; is the sample data after disturbance; The perturbed sample data is predicted by the model as class probability; The category output by the model; For data categories.

[0044] Preferably, the dynamic distillation loss function is used to measure the difference between the new and old knowledge of the model to dynamically adjust the loss function weight; its expression is:

[0045] ;

[0046] in, is the dynamic distillation loss function; T is the task set; t is the current task; For the model; is the category prediction model; is the i-th sample data; is one-hot encoding; The prediction results output by the model; for The corresponding prediction results; is the model prediction probability corresponding to the i-th sample data; Indicates a trained task.

[0047] The present invention also provides a time series data incremental learning and prediction device, comprising:

[0048] An acquisition unit, used to acquire multi-dimensional time series data of industrial equipment with multi-source and heterogeneous distribution;

[0049] A time-frequency feature alignment unit is used to perform sliding window segmentation and feature alignment on the multidimensional time series data, and then convert it into a two-dimensional spectrum through Mel spectrum analysis;

[0050] An incremental learning unit is configured to input the two-dimensional spectrogram into a pre-trained basic model for incremental learning training, fix the key parameters of the basic model's historical tasks based on an elastic weight solidification module, and constrain the parameter updates of the basic model by adding regularization terms to reduce deviations from important parameters of previous tasks;

[0051] The experience replay unit is used to calculate the representativeness and uncertainty of samples in real time during the incremental learning training process, screen out representative samples and complex samples, and save them to the replay buffer for sample replay during the next training;

[0052] A dynamic distillation unit, configured to combine current task data with replay samples in the replay buffer, and based on a dynamic distillation loss function, achieve fusion of new and old knowledge through inter-layer feature distillation to obtain an incremental learning prediction model;

[0053] The prediction unit is used to process the newly acquired multi-dimensional time series data of industrial equipment into a two-dimensional spectrum diagram and then input the two-dimensional spectrum diagram into the incremental learning prediction model for prediction to obtain the prediction result of the operating status of the industrial equipment.

[0054] The present invention also provides a time series data incremental learning and prediction device, including a processor and a memory, wherein the memory stores a computer program, and the computer program can be executed by the processor to implement a time series data incremental learning and prediction method as described above.

[0055] The present invention also provides a computer-readable storage medium, on which computer-readable instructions are stored. When the computer-readable instructions are executed by a processor of a device where the computer-readable storage medium is located, the method for incremental learning and prediction of time series data as described above is implemented.

[0056] In summary, compared with the prior art, the present invention has the following beneficial effects:

[0057] The present invention uses Mel spectrum analysis to convert the original time series data into a two-dimensional spectrum graph suitable for deep learning model processing, which retains the time series characteristics while improving the input quality of the model; in the incremental learning process, the elastic weight solidification module and regularization constraints are used to significantly reduce the model's forgetting of historical task knowledge; through the dual screening mechanism of representative samples and complex samples, the diversity and representativeness of samples in the playback buffer are ensured; through the improved dynamic distillation loss function, the model is explicitly guided to learn the knowledge association between new and old tasks, thereby improving the model's generalization ability and stability.

[0058] The present invention solves the problem of catastrophic forgetting in dynamic scenarios of massive industrial time series data through a multi-strategy collaborative incremental learning mechanism, significantly improves the continuous learning accuracy and stability of the model in tasks such as equipment status monitoring and fault warning, and has important practical application value. BRIEF DESCRIPTION OF THE DRAWINGS

[0059] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the embodiments. It should be understood that the following drawings only illustrate certain embodiments of the present invention and therefore should not be regarded as limiting the scope. For ordinary technicians in this field, other relevant drawings can be obtained based on these drawings without paying any creative work.

[0060] Figure 1 A flowchart of a time series data incremental learning prediction method provided in Example 1.

[0061] Figure 2 A framework diagram of a time series data incremental learning prediction method provided in Example 1.

[0062] Figure 3 A schematic diagram of a time series data incremental learning and prediction device provided in Example 2.

[0063] The present invention is further described in detail below with reference to the accompanying drawings and specific embodiments. DETAILED DESCRIPTION

[0064] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention. Therefore, the following detailed description of the embodiments of the present invention provided in the drawings is not intended to limit the scope of the invention for which protection is sought, but merely represents selected embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0065] Example 1

[0066] Embodiment 1 of the present invention provides a method for incremental learning and prediction of time series data, which can be implemented by a time series data incremental learning and prediction device (hereinafter referred to as the prediction device), and in particular, executed by one or more processors in the prediction device.

[0067] In this embodiment, the prediction device may be an electronic device equipped with a processor, which has a computer program for the time series data incremental learning prediction method and can be executed, such as a computer, a smart phone, a smart tablet, a workstation, etc., which is not limited here.

[0068] like Figure 1-Figure 2 As shown, a time series data incremental learning prediction method includes steps S1 to S6.

[0069] S1, obtains multi-dimensional time series data of industrial equipment with multi-source and heterogeneous distribution.

[0070] In this step, we first identify the data sources, which include sensor data from industrial equipment, control system data, and external data.

[0071] Sensor data: Industrial equipment is typically equipped with a variety of sensors, such as temperature, pressure, and vibration sensors. These sensors continuously collect various parameters during equipment operation, generating time-series data. For example, in a wind turbine, temperature sensors record temperature changes in various components, while pressure sensors monitor pressure fluctuations in the hydraulic system.

[0072] Control system data: The control system of industrial equipment records information such as the equipment's operating status and control instructions. This data, also stored in a time series format, reflects the equipment's operational processes and operating logic. For example, on an automated production line, a PLC (programmable logic controller) records motor start and stop times, as well as parameter settings for each production step.

[0073] External data: In addition to the device's own data, consider acquiring external data related to the device's operation, such as ambient temperature, humidity, wind speed, etc. This data may affect the performance and operating status of the device, and taking it into consideration can improve the accuracy of the prediction.

[0074] Next, data collection and storage are performed. A dedicated data acquisition system can be established to ensure real-time and accurate data acquisition from various data sources. This acquisition system needs to be highly reliable and stable to withstand the complex conditions of an industrial environment. For example, an industrial-grade data acquisition card can be used to connect to sensors and control systems using appropriate communication protocols (such as Modbus and OPC UA).

[0075] The collected data is then stored in an appropriate database for subsequent processing and analysis. You can choose a relational database (such as MySQL or Oracle) to store structured data, or a time series database (such as InfluxDB or TimescaleDB) to efficiently store and query time series data. Also, consider the data storage cycle and backup strategy to prevent data loss.

[0076] Then, the collected multi-source heterogeneous data is preprocessed, including data cleaning and data fusion.

[0077] Data cleaning: Due to noise and interference in industrial environments, collected data may contain missing values, outliers, and other issues. Data cleaning is necessary to fill missing values ​​(e.g., using the mean, median, or interpolation methods) and identify and correct outliers (e.g., using statistical methods or machine learning algorithms).

[0078] Data fusion: Data from multiple sources may have different formats, units, and sampling frequencies. Data fusion is necessary to unify them under the same framework. For example, for data with different sampling frequencies, resampling methods (such as upsampling and downsampling) can be used to bring them to the same sampling rate; for data with different units, unit conversion can be performed.

[0079] S2, after performing sliding window segmentation and feature alignment on the multidimensional time series data, converting it into a two-dimensional spectrum graph through Mel spectrum analysis.

[0080] The parameters of the Mel spectrum include: hop length, window length, window type, and number of Mel filters. The parameters of the Mel spectrum need to be determined according to the application scenario, such as hop length of 160, window length of 512, window type of Hanning window, and number of Mel filters of 48.

[0081] The specific steps of S2 are:

[0082] S21, dividing the multi-dimensional time series data according to the preset jump length and window length. Each sample divided is a time series segment of fixed length, containing data of multiple dimensions.

[0083] S22, perform feature alignment on the segmented data, including time alignment and feature dimension alignment.

[0084] Time alignment: Data in different dimensions may have different sampling frequencies or collection time points. Interpolation methods (such as linear interpolation) are used to adjust the time-mismatched data so that they have the same timestamp.

[0085] Feature dimension alignment: If the sample feature dimensions are inconsistent, you can handle it by filling missing values ​​(such as with 0 or specific markers) or deleting incomplete samples.

[0086] S23, based on the set window type (such as Hanning window), performs Fourier transform on the data after feature alignment in each window, converting it from the time domain to the frequency domain, and obtains the linear spectrum, that is, the energy distribution of the signal at different times and frequencies.

[0087] S24 , filtering the linear spectrum through a set number of Mel filters to obtain a Mel spectrum.

[0088] Mel filter banks can simulate the human ear's perception of sounds of different frequencies.

[0089] S25, performing a logarithmic operation on the Mel spectrum to compress the dynamic range of the spectrum and make the features more obvious, obtaining a logarithmic Mel spectrum, and visualizing it as a two-dimensional spectrum graph.

[0090] S3, input the two-dimensional spectrum graph into the pre-trained basic model for incremental learning training, and fix the historical task key parameters of the basic model based on the elastic weight solidification module, and constrain the parameter update of the basic model by adding regularization terms to reduce the deviation from the important parameters of the previous task.

[0091] Before the two-dimensional spectrogram is fed into the base model, rows containing null values ​​are filtered to improve model training quality and prevent model error accumulation. Null value rows can make the matrix structure incomplete. If directly fed into the model, the model will attempt to process this meaningless data, resulting in incorrect feature representations.

[0092] The basic model is a deep learning model with hierarchical time-frequency window self-attention, including a partitioning module, a position embedding module, a hierarchical multi-head self-attention structure module and a window semantic module; and the basic model is trained using relevant data sets to enable it to learn universal feature representations.

[0093] The division module is used to divide the two-dimensional spectrum into equally spaced windows, and each window is divided into a number of time-frequency blocks;

[0094] The position embedding module is used to add position embedding information to each time-frequency block;

[0095] The hierarchical multi-head self-attention structure module is used to perform self-attention calculation on the time-frequency blocks after position embedding in each layer of the multi-head self-attention structure, and perform time-frequency dimension compression and channel dimension expansion while extracting the relationship between blocks;

[0096] The window semantic module is used to map the data after channel dimension expansion to the category semantic dimension through single-layer convolution and full connection.

[0097] In this embodiment, the category semantic dimension refers to the label category dimension of the current data set.

[0098] For example, the two-dimensional spectrogram (assuming a size of 256×256, with a time frame dimension of 256 and a frequency dimension of 256) is divided into eight equally spaced windows along the time frame and frequency dimensions. This is then divided into 4×4 time-frequency blocks (each block is 8×64) with equal spacing along the time frame and frequency dimensions. A position embedding is added to each time-frequency block to provide position information. Four layers of multi-head self-attention structures are stacked, each layer compressing the frequency and time dimensions by 1 / 2 (via average pooling or linear projection) and expanding the channel dimension by a factor of 4. The expanded channel features are then mapped to the category semantic dimension (assuming the number of categories is 10, the channel dimension of the expanded features can be mapped to 10) through a single convolutional layer and a fully connected layer in the window semantic module to obtain a feature embedding. The number of self-attention blocks within each of the four-layer multi-head self-attention structures in this example can be set to 4, 8, 8, or 4, or as required, and is not limited here.

[0099] The elastic weight solidification module is used to fix the historical task key parameters of the basic model to prevent catastrophic forgetting in incremental learning. Specifically:

[0100] After each task training is completed, the importance score of the parameters in the basic model to the task is calculated through the Fisher information matrix. The calculation formula is:

[0101] ;

[0102] in, Score the importance of the t-th task; is the learning training data for the tth task; is the i-th parameter of the basic model; For the input sample data of the basic model Output category label of For parameters The basic model of is the true category label; Indicates derivative.

[0103] In this embodiment, the Fisher Information Matrix (FIM) is an important concept in statistics and machine learning, and is defined as the expected negative value of the Hessian matrix (second-order derivative matrix) of the log-likelihood function.

[0104] In addition, regularization terms are added to the model's objective function to constrain the parameter updates of the base model. For example, L2 regularization terms are used to keep the model parameters as small as possible during the update process, reducing the deviation from the important parameters of the previous task.

[0105] S4, during the incremental learning training process, calculates the representativeness and uncertainty of the samples in real time, screens out representative samples and complex samples, and saves them to the playback buffer for sample playback during the next training.

[0106] When calculating the representativeness of a sample, this embodiment expresses it by calculating the distance between the sample and the class prototype of the task category. The farther the distance, the stronger the representativeness. The class prototype distance is calculated as follows:

[0107] First, for each category in each task, the feature mean of the data in the task is calculated as the class prototype of the category. The formula is:

[0108] ;

[0109] in, For category The characteristic mean of For category The sample set of is the sample data, is a feature extractor based on the basic model;

[0110] Calculate the cosine similarity of all samples to the class prototype. The formula is:

[0111] ;

[0112] in, for arrive The cosine similarity of Represents sample data The characteristic mean of .

[0113] When screening complex samples, the uncertainty of the sample data is calculated based on the sample representativeness calculation results for screening, as follows:

[0114] First, different data perturbations are applied to the input training data, and the model prediction results are obtained based on the perturbed data;

[0115] The uncertainty of the sample data is calculated by the stability of the prediction results, and complex samples are screened out. For example, a threshold can be set to classify samples with uncertainty higher than the threshold as complex samples. The expression is:

[0116] ;

[0117] in, is the sample data; Task collection; For the current task; is the sample data after disturbance; The perturbed sample data is predicted by the model as class probability; The category output by the model; For data categories.

[0118] The selected representative samples and complex samples are saved in the playback buffer for sample playback during the next training.

[0119] S5, combining the current task data with the playback samples in the playback buffer, based on the dynamic distillation loss function, realizes the fusion of new and old knowledge through inter-layer feature distillation to obtain an incremental learning prediction model.

[0120] The dynamic distillation loss function is used to measure the difference between the new and old knowledge of the model to dynamically adjust the loss function weight; its expression is:

[0121] ;

[0122] in, is the dynamic distillation loss function; T is the task set; t is the current task; For the model; is the category prediction model; is the i-th sample data; is one-hot encoding; The category output by the model; for The corresponding prediction results; is the model prediction probability corresponding to the i-th sample data; Indicates a trained task.

[0123] During model training, feature representations are extracted from different layers. For example, for a CNN model, features are extracted from the convolutional, pooling, and fully connected layers. Using a dynamic distillation loss function, the inter-layer features of the model on the new task are aligned as closely as possible with those on the old task, enabling the fusion of new and old knowledge.

[0124] The model is trained using a dynamic distillation loss function, combining the current task data with the replay samples in the replay buffer. During training, the model parameters are continuously adjusted to achieve good performance on both new and old tasks.

[0125] According to the training results, the model parameters are updated to obtain an incremental learning prediction model for subsequent tasks.

[0126] S6, processing the newly acquired multi-dimensional time series data of the industrial equipment into a two-dimensional spectrum graph and inputting the two-dimensional spectrum graph into the incremental learning prediction model for prediction, thereby obtaining a prediction result of the operating status of the industrial equipment.

[0127] In this step, the newly acquired multidimensional time series data of industrial equipment is processed in the same way as steps S1 and S2, including data cleaning, fusion, sliding window segmentation, feature alignment, and Mel spectrum analysis to generate a two-dimensional spectrum graph.

[0128] The generated two-dimensional spectrogram is fed into an incremental learning prediction model. The model processes the input two-dimensional spectrogram and outputs a prediction of the future operating state of the industrial equipment. For example, in a device fault diagnosis task, the model outputs a prediction of whether the equipment will fail and the type of failure.

[0129] Evaluate the prediction results, such as calculating accuracy, recall, and other indicators. Based on the prediction results, take appropriate measures, such as performing equipment maintenance in advance, to avoid equipment failure and production losses.

[0130] In the actual equipment status monitoring task, the original time series data is first preprocessed through sliding window segmentation and time-frequency feature alignment to generate a high-quality two-dimensional spectrum graph; then, the model is built and trained through incremental learning to ensure that the model retains historical task knowledge during the new task learning process; then, representative samples and complex samples are screened and added to the playback buffer; finally, the improved dynamic distillation loss function is integrated to achieve the effective fusion of new and old knowledge.

[0131] In summary, compared with the prior art, the present invention has the following beneficial effects:

[0132] The present invention uses Mel spectrum analysis to convert the original time series data into a two-dimensional spectrum graph suitable for deep learning model processing, which retains the time series characteristics while improving the input quality of the model; in the incremental learning process, the elastic weight solidification module and regularization constraints are used to significantly reduce the model's forgetting of historical task knowledge; through the dual screening mechanism of representative samples and complex samples, the diversity and representativeness of samples in the playback buffer are ensured; through the improved dynamic distillation loss function, the model is explicitly guided to learn the knowledge association between new and old tasks, thereby improving the model's generalization ability and stability.

[0133] The present invention solves the problem of catastrophic forgetting in dynamic scenarios of massive industrial time series data through a multi-strategy collaborative incremental learning mechanism, significantly improves the continuous learning accuracy and stability of the model in tasks such as equipment status monitoring and fault warning, and has important practical application value.

[0134] Example 2

[0135] like Figure 3 As shown, the second embodiment of the present invention further provides a time series data incremental learning and prediction device, comprising:

[0136] An acquisition unit, used to acquire multi-dimensional time series data of industrial equipment with multi-source and heterogeneous distribution;

[0137] A time-frequency feature alignment unit is used to perform sliding window segmentation and feature alignment on the multidimensional time series data, and then convert it into a two-dimensional spectrum through Mel spectrum analysis;

[0138] An incremental learning unit is configured to input the two-dimensional spectrogram into a pre-trained basic model for incremental learning training, fix the key parameters of the basic model's historical tasks based on an elastic weight solidification module, and constrain the parameter updates of the basic model by adding regularization terms to reduce deviations from important parameters of previous tasks;

[0139] The experience replay unit is used to calculate the representativeness and uncertainty of samples in real time during the incremental learning training process, screen out representative samples and complex samples, and save them to the replay buffer for sample replay during the next training;

[0140] A dynamic distillation unit, configured to combine current task data with replay samples in the replay buffer, and based on a dynamic distillation loss function, achieve fusion of new and old knowledge through inter-layer feature distillation to obtain an incremental learning prediction model;

[0141] The prediction unit is used to process the newly acquired multi-dimensional time series data of industrial equipment into a two-dimensional spectrum diagram and then input the two-dimensional spectrum diagram into the incremental learning prediction model for prediction to obtain the prediction result of the operating status of the industrial equipment.

[0142] Example 3

[0143] The third embodiment of the present invention also provides a time series data incremental learning prediction device, which includes a memory and a processor, wherein the memory stores a computer program, and the computer program can be executed by the processor to implement the time series data incremental learning prediction method as described above.

[0144] Example 4

[0145] The fourth embodiment of the present invention also provides a computer-readable storage medium, which stores computer-readable instructions. When the computer-readable instructions are executed by the processor of the device where the computer-readable storage medium is located, the incremental learning and prediction method for time series data as described above is implemented.

[0146] In the several embodiments provided in the embodiments of the present invention, it should be understood that the disclosed devices and methods can also be implemented in other ways. The device and method embodiments described above are merely illustrative. For example, the flowcharts in the accompanying drawings show the possible architectures, functions, and operations of the devices, methods, and computer program products according to multiple embodiments of the present invention. In this regard, each box in the flowchart or block diagram can represent a module, program segment, or part of a code, which contains one or more executable instructions for implementing the specified logical functions. It should also be noted that in some alternative implementations, the functions marked in the boxes can also occur in an order different from that marked in the drawings. For example, two consecutive boxes can actually be executed substantially in parallel, and they can sometimes be executed in the opposite order, depending on the functions involved. It should also be noted that each box in the block diagram and / or flowchart, as well as the combination of boxes in the block diagram and / or flowchart, can be implemented using a dedicated hardware-based system that performs the specified functions or actions, or can be implemented using a combination of dedicated hardware and computer instructions.

[0147] In addition, the functional modules in the various embodiments of the present invention may be integrated together to form an independent part, or each module may exist independently, or two or more modules may be integrated to form an independent part.

[0148] If the functions are implemented in the form of software modules and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the portion that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions for causing a computer device (which can be a personal computer, electronic device, or network device, etc.) to perform all or part of the steps of the methods described in various embodiments of the present invention. The aforementioned storage media include various media that can store program code, such as USB flash drives, mobile hard drives, read-only memories (ROMs), random access memories (RAMs), magnetic disks, or optical disks. It should be noted that, in this document, the terms "comprise," "include," or any other variations thereof are intended to encompass non-exclusive inclusion, such that a process, method, article, or device that includes 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. Without further constraints, an element defined by the phrase "comprises a..." does not preclude the existence of additional identical elements in the process, method, article or apparatus that includes the element.

[0149] The terms used in the embodiments of the present invention are only for the purpose of describing specific embodiments and are not intended to limit the present invention. The singular forms "a", "an", "the" and "the" used in the embodiments of the present invention and the appended claims are also intended to include plural forms unless the context clearly indicates otherwise.

[0150] It should be understood that the term "and / or" as used herein is merely a description of the relationship between associated objects, indicating that three possible relationships exist. For example, "A and / or B" can represent: A exists alone, A and B exist simultaneously, or B exists alone. Furthermore, the character " / " in this document generally indicates that the associated objects are in an "or" relationship.

[0151] The word "if," as used herein, may be interpreted as "at the time of" or "when" or "in response to determining" or "in response to detecting," depending on the context. Similarly, the phrases "if it is determined" or "if (stated condition or event) is detected" may be interpreted as "when it is determined" or "in response to the determination" or "when detecting (stated condition or event)" or "in response to detecting (stated condition or event)," depending on the context.

[0152] The "first" and "second" mentioned in the embodiments are merely used to distinguish similar objects and do not represent a specific ordering of the objects. It is understood that the specific order or precedence of "first" and "second" can be interchanged where appropriate. It should be understood that the objects distinguished by "first" and "second" can be interchanged where appropriate, so that the embodiments described herein can be implemented in an order other than that illustrated or described herein.

[0153] The foregoing description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Those skilled in the art will readily appreciate that various modifications and variations of the present invention are possible. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present invention are intended to be within the scope of protection of the present invention.

Claims

1. A time series data incremental learning prediction method, characterized in that: include: S1, obtains multi-dimensional time series data of industrial equipment with multi-source and heterogeneous distribution; S2, performing sliding window segmentation and feature alignment on the multidimensional time series data, and converting it into a two-dimensional spectrogram through Mel spectrum analysis; S3, inputting the two-dimensional spectrogram into a pre-trained basic model for incremental learning training, and fixing the key parameters of the historical tasks of the basic model based on the elastic weight solidification module, and constraining the parameter update of the basic model by adding a regularization term to reduce the deviation from the important parameters of the previous task; wherein, the elastic weight solidification module is used to fix the key parameters of the historical tasks of the basic model to prevent catastrophic forgetting in incremental learning, specifically: After each task training is completed, the importance score of the parameters in the basic model to the task is calculated through the Fisher information matrix. The calculation formula is: ； in, Score the importance of the t-th task; is the learning training data for the tth task; is the i-th parameter of the basic model; For the input sample data of the basic model Output category label of For parameters The basic model of is the true category label; represents the derivation; S4, during the incremental learning training process, calculates the representativeness and uncertainty of the samples in real time, screens out representative samples and complex samples, and saves them to the playback buffer for sample playback during the next training. When calculating the uncertainty of the samples in real time and screening out complex samples: First, different data perturbations are applied to the input training data, and the model prediction results are obtained based on the perturbed data; The uncertainty of sample data is calculated by the stability of the prediction results, and complex samples are screened out; uncertainty The expression is: ； in, is the sample data; Task collection; For the current task; is the sample data after disturbance; The perturbed sample data is predicted by the model as class probability; The category output by the model; is the data category; S5, combining the current task data with the playback samples in the playback buffer, based on the dynamic distillation loss function, realizes the fusion of new and old knowledge through inter-layer feature distillation to obtain an incremental learning prediction model; S6, processing the newly acquired multi-dimensional time series data of the industrial equipment into a two-dimensional spectrum graph and inputting the two-dimensional spectrum graph into the incremental learning prediction model for prediction, thereby obtaining a prediction result of the operating status of the industrial equipment.

2. A time series data incremental learning prediction method according to claim 1, characterized in that ,The parameters of the Mel spectrum include: jump length, window length, window type and number of Mel filters; The S2 is specifically: Segmenting the multidimensional time series data according to a preset jump length and window length; Perform feature alignment on the segmented data; Based on the set window type, the data after feature alignment in each window is Fourier transformed to convert it from the time domain to the frequency domain to obtain a linear spectrum; Filtering the linear spectrum through a set number of Mel filters to obtain a Mel spectrum; A logarithmic operation is performed on the Mel spectrum to compress the dynamic range of the spectrum, thereby obtaining a logarithmic Mel spectrum, which is visualized as a two-dimensional spectrum graph.

3. The time series data incremental learning prediction method according to claim 1 is characterized in that , further comprising: filtering the rows of null values ​​in the two-dimensional spectrum graph before the two-dimensional spectrum graph is input into the basic model; The basic model is a deep learning model with hierarchical time-frequency window self-attention, including a partitioning module, a position embedding module, a hierarchical multi-head self-attention structure module and a window semantic module; The division module is used to divide the two-dimensional spectrum into equally spaced windows, and each window is divided into a number of time-frequency blocks; The position embedding module is used to add position embedding information to each time-frequency block; The hierarchical multi-head self-attention structure module is used to perform self-attention calculation on the time-frequency blocks after position embedding in each layer of the multi-head self-attention structure, and perform time-frequency dimension compression and channel dimension expansion while extracting the relationship between blocks; The window semantic module is used to map the data after channel dimension expansion to the category semantic dimension through single-layer convolution and full connection.

4. The time series data incremental learning prediction method according to claim 1 is characterized in that When calculating the representativeness of a sample, it is reflected by the distance between the sample and the prototype of the task category. The calculation method of the prototype distance is: First, for each category in each task, the feature mean of the data in the task is calculated as the class prototype of the category. The formula is: ； in, For category The characteristic mean of For category The sample set of is the sample data, is a feature extractor based on the basic model; Calculate the cosine similarity of all samples to the class prototype. The formula is: ； in, for arrive The cosine similarity of Represents sample data The characteristic mean of .

5. The time series data incremental learning prediction method according to claim 1 is characterized in that ,The dynamic distillation loss function is used to measure the difference between the model and the old knowledge to dynamically adjust the loss function weight; its expression is : ； in, is the dynamic distillation loss function; T is the task set; t is the current task; For the model; is the category prediction model; is the i-th sample data; is one-hot encoding; The category output by the model; for The corresponding prediction results; is the model prediction probability corresponding to the i-th sample data; Indicates a trained task.

6. A time series data incremental learning prediction device, characterized in that: include: An acquisition unit, used to acquire multi-dimensional time series data of industrial equipment with multi-source and heterogeneous distribution; A time-frequency feature alignment unit is used to perform sliding window segmentation and feature alignment on the multidimensional time series data, and then convert it into a two-dimensional spectrum through Mel spectrum analysis; An incremental learning unit is configured to input the two-dimensional spectrogram into a pre-trained basic model for incremental learning training, and fix the key parameters of the basic model's historical tasks based on an elastic weight solidification module, and constrain the parameter updates of the basic model by adding regularization terms to reduce deviations from important parameters of previous tasks. The elastic weight solidification module is configured to fix the key parameters of the basic model's historical tasks to prevent catastrophic forgetting in incremental learning, specifically: After each task training is completed, the importance score of the parameters in the basic model to the task is calculated through the Fisher information matrix. The calculation formula is: ； in, Score the importance of the t-th task; is the learning training data for the tth task; is the i-th parameter of the basic model; For the input sample data of the basic model Output category label of For parameters The basic model of is the true category label; represents the derivation; The experience replay unit is used to calculate the representativeness and uncertainty of samples in real time during incremental learning training, filter out representative samples and complex samples, and save them to the replay buffer for sample replay during the next training. When calculating the uncertainty of samples in real time and filtering out complex samples: First, different data perturbations are applied to the input training data, and the model prediction results are obtained based on the perturbed data; The uncertainty of sample data is calculated by the stability of the prediction results, and complex samples are screened out; uncertainty The expression is: ； in, is the sample data; Task collection; For the current task; is the sample data after disturbance; The perturbed sample data is predicted by the model as class probability; The category output by the model; is the data category; A dynamic distillation unit, configured to combine current task data with replay samples in the replay buffer, and based on a dynamic distillation loss function, achieve fusion of new and old knowledge through inter-layer feature distillation to obtain an incremental learning prediction model; The prediction unit is used to process the newly acquired multi-dimensional time series data of industrial equipment into a two-dimensional spectrum diagram and then input the two-dimensional spectrum diagram into the incremental learning prediction model for prediction to obtain the prediction result of the operating status of the industrial equipment.

7. A time series data incremental learning prediction device, characterized in that: It includes a processor and a memory, wherein the memory stores a computer program, and the computer program can be executed by the processor to implement a time series data incremental learning and prediction method as described in any one of claims 1 to 5.

8. A computer-readable storage medium, characterized in that The computer-readable storage medium stores computer-readable instructions, and when the computer-readable instructions are executed by the processor of the device where the computer-readable storage medium is located, the time series data incremental learning and prediction method according to any one of claims 1 to 5 is implemented.

Images