Real-time feature extraction method and system for built-in timing data of intelligent converged terminal

Abstract

The application discloses a real-time feature extraction method and system for built-in time sequence data of an intelligent fusion terminal, relates to the technical field of data processing, and comprises the following steps: obtaining an original time sequence data stream through the intelligent fusion terminal, performing multi-modal data fusion, and constructing a fusion time sequence data stream; constructing a double-layer feature extraction window, performing real-time feature fusion extraction on the fusion time sequence data stream, and obtaining a multi-scale real-time feature vector; constructing an initial feature vector matrix based on the multi-scale real-time feature vector, performing dynamic feature weight distribution updating, and obtaining a feature vector matrix; performing state recognition response on the intelligent fusion terminal according to the feature vector matrix, and generating an auxiliary decision instruction of the intelligent fusion terminal. The application solves the technical problem that the real-time feature extraction capability of the intelligent fusion terminal for built-in time sequence data is insufficient in the prior art, and achieves the technical effect of improving the real-time feature extraction effect of the intelligent fusion terminal for built-in time sequence data.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, specifically to a method and system for real-time feature extraction of time-series data built into intelligent fusion terminals. Background Technology

[0002] As the application of intelligent converged terminals in target application scenarios continues to increase, the built-in sensing units of the terminals will continuously generate a large amount of time-series data during operation. This type of data is often characterized by multi-source heterogeneity, continuous change and strong real-time requirements. If there is a lack of targeted real-time processing mechanisms, it is easy to lead to insufficient data fusion and delayed extraction of key features, making it difficult to form stable and effective feature expressions in a timely manner. As a result, the intelligent converged terminal has insufficient real-time feature extraction capabilities for built-in time-series data. Summary of the Invention

[0003] This application provides a method and system for real-time feature extraction of built-in time-series data in intelligent fusion terminals, which is used to address the technical problem that existing intelligent fusion terminals have insufficient real-time feature extraction capabilities for built-in time-series data.

[0004] In view of the above problems, this application provides a method and system for real-time feature extraction of time series data built into intelligent fusion terminals.

[0005] A first aspect of this application provides a method for real-time feature extraction of time-series data built into an intelligent fusion terminal, the method comprising:

[0006] The intelligent fusion terminal performs real-time data acquisition and sensing of the target application scenario to obtain the raw time-series data stream; multimodal data fusion is performed on the raw time-series data stream to construct a fused time-series data stream; a two-layer feature extraction window is constructed, the two-layer feature extraction window mechanism including a decay window and a dynamic sliding window, and real-time feature fusion extraction is performed on the fused time-series data stream through the decay window and the dynamic sliding window to obtain multi-scale real-time feature vectors; an initial feature vector matrix is ​​constructed based on the multi-scale real-time feature vectors, and dynamic feature weight allocation and updating are performed to obtain a feature vector matrix; the intelligent fusion terminal performs state recognition response based on the feature vector matrix to generate auxiliary decision-making instructions for the intelligent fusion terminal.

[0007] A second aspect of this application provides a real-time feature extraction system for time-series data built into an intelligent fusion terminal, the system comprising: The system includes a data acquisition module for real-time acquisition and sensing of the target application scenario via an intelligent fusion terminal to obtain raw time-series data streams; a data fusion module for multimodal data fusion of the raw time-series data streams to construct a fused time-series data stream; a feature fusion and extraction module for constructing a two-layer feature extraction window, which includes a decay window and a dynamic sliding window, to perform real-time feature fusion extraction on the fused time-series data stream and obtain multi-scale real-time feature vectors; a weight allocation and update module for constructing an initial feature vector matrix based on the multi-scale real-time feature vectors and performing dynamic feature weight allocation and update to obtain a feature vector matrix; and an instruction generation module for responding to the intelligent fusion terminal's state recognition based on the feature vector matrix and generating auxiliary decision-making instructions for the intelligent fusion terminal.

[0008] One or more technical solutions provided in this application have at least the following technical effects or advantages: This application utilizes an intelligent fusion terminal to perform real-time data acquisition and sensing of a target application scenario, obtaining a raw time-series data stream. Multimodal data fusion is then performed on the raw time-series data stream to construct a fused time-series data stream. A dual-layer feature extraction window is constructed, comprising a decay window and a dynamic sliding window mechanism. Real-time feature fusion extraction is performed on the fused time-series data stream using the decay window and the dynamic sliding window to obtain multi-scale real-time feature vectors. Based on the multi-scale real-time feature vectors, an initial feature vector matrix is ​​constructed and dynamically updated with feature weight allocation to obtain a feature vector matrix. The intelligent fusion terminal responds to the feature vector matrix with state recognition, generating auxiliary decision-making instructions for the intelligent fusion terminal. This invention addresses the technical problem of insufficient real-time feature extraction capability of existing intelligent fusion terminals for built-in time-series data. By constructing a dual-layer feature extraction window including a decay window and a dynamic sliding window, real-time feature fusion extraction is performed on the fused time-series data stream, thereby improving the real-time feature extraction effect of the intelligent fusion terminal for built-in time-series data. Attached Figure Description

[0009] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0010] Figure 1 A schematic diagram of the real-time feature extraction method for built-in time series data in an intelligent fusion terminal provided in this application embodiment; Figure 2A schematic diagram of the structure of a real-time feature extraction system for built-in time-series data in an intelligent fusion terminal provided in this application embodiment.

[0011] Figure labeling: Data acquisition module 11, data fusion module 12, feature fusion and extraction module 13, weight allocation and update module 14, instruction generation module 15. Detailed Implementation

[0012] This application provides a method and system for real-time feature extraction of built-in time-series data in intelligent fusion terminals. It addresses the technical problem of insufficient real-time feature extraction capability of built-in time-series data in existing technologies by constructing a two-layer feature extraction window including a decay window and a dynamic sliding window to perform real-time feature fusion extraction of the fused time-series data stream, thereby improving the real-time feature extraction effect of built-in time-series data in intelligent fusion terminals.

[0013] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the scope of protection of this application.

[0014] It should be noted that any variation of the terms "comprising" and "having" is intended to cover non-exclusive inclusion, for example, a process, method, system, product, or server that includes a series of steps or units is not necessarily limited to those steps or units that are explicitly listed, but may include other steps or modules that are not explicitly listed or that are inherent to such processes, methods, products, or devices.

[0015] Example 1, as Figure 1 As shown, this application provides a real-time feature extraction method for time-series data built into an intelligent fusion terminal, the method comprising: Step S100: Real-time data collection and sensing of the target application scenario is performed through the intelligent fusion terminal to obtain the raw time-series data stream.

[0016] In this embodiment, the intelligent fusion terminal is deployed in the corresponding target application scenario, such as equipment operation monitoring, environmental change perception, or personnel status recognition. The sensors built into the intelligent fusion terminal continuously collect status information in the target application scenario, obtaining multiple sets of sensor data corresponding to different sampling times. Here, the target application scenario refers to the actual application environment that requires status perception and data acquisition, and real-time sensor acquisition refers to continuously executing the data acquisition process in a continuous time sequence. During the acquisition process, the intelligent fusion terminal matches and records the sensor data acquired at each moment with the corresponding acquisition time, and then arranges the sensor data at each moment continuously in chronological order, thereby forming a data set that unfolds sequentially over time, and finally obtaining the original time-series data stream.

[0017] Step S200: Perform multimodal data fusion on the original time-series data stream to construct a fused time-series data stream.

[0018] In this embodiment, when performing multimodal data fusion on the original time-series data stream, a reference time axis is first constructed, and deviation analysis is performed on the original time-series data stream based on the reference time axis to extract time offset data. Then, the original time-series data stream is synchronized and aligned according to the time offset data and the reference time axis to generate a time-aligned dataset. Subsequently, local data analysis is performed on the time-aligned dataset to obtain local data distribution characteristics, and instantaneous abnormal fluctuation analysis is performed based on the local data distribution characteristics to extract instantaneous abnormal jump points. Then, the time-aligned dataset is cleaned according to the instantaneous abnormal jump points to obtain a cleaned time-series dataset. Finally, the cleaned time-series dataset is traversed to determine multiple data gap positions, and adjacent searches are performed according to the multiple data gap positions to extract multiple valid sensor data points to fill the multiple data gap positions, thereby generating a fused time-series data stream.

[0019] Furthermore, in the method provided in the application embodiments, performing multimodal data fusion on the original time-series data stream to construct a fused time-series data stream further includes: A reference time axis is constructed, and the original time-series data stream is subjected to deviation analysis according to the reference time axis to extract time offset data. The original time-series data stream is updated according to the time offset data and synchronized with the reference time axis to generate a time-series aligned dataset. The time-series aligned dataset is traversed to perform local data analysis to obtain local data distribution characteristics. Based on the local data distribution characteristics, instantaneous abnormal fluctuation analysis is performed to extract instantaneous abnormal jump points. The time-series aligned dataset is cleaned according to the instantaneous abnormal jump points to obtain a cleaned time-series dataset. The cleaned time-series dataset is traversed to mark gaps and determine multiple data gap positions. Based on the cleaned time-series dataset, an adjacency search is performed according to the multiple data gap positions to extract multiple valid sensor data points. The multiple valid sensor data points are filled into the multiple data gap positions to generate the fused time-series data stream.

[0020] In this embodiment, a reference time axis is first constructed, and the original time-series data stream is subjected to deviation analysis according to the reference time axis to extract time offset data. Specifically, the acquisition time information corresponding to each data point in the original time-series data stream is first read, and a reference time axis is constructed according to a unified time order. The reference time axis is a unified time reference used for time comparison and time correction of each data point in the original time-series data stream. Then, the acquisition time corresponding to each data point in the original time-series data stream is compared with the corresponding time position on the reference time axis to determine the difference between the current time position of each data point and the corresponding standard time position. Time offset data is extracted based on the difference. The time offset data is used to characterize the deviation of each data point relative to the reference time axis, which indicates a time shift forward or backward.

[0021] Next, the original time-series data stream is updated according to the time offset data, and synchronized with the reference time axis to generate a time-aligned dataset. Specifically, based on the time offset data corresponding to each data point, the time stamps of each data point in the original time-series data stream are corrected so that the time position of each data point is adjusted to the corresponding standard time position on the reference time axis. After the time position update is completed, the updated data points are synchronized with the reference time axis. Synchronization alignment means unifying the time position of each data point according to a unified time base, so that each data point establishes a time correspondence under the same time base. Finally, the synchronized data points are organized according to the time order of the reference time axis to generate a time-aligned dataset.

[0022] Subsequently, local data analysis is performed by traversing the time-series aligned dataset. This process begins with a local sliding analysis based on the time series, defining a local neighborhood analysis window and determining the first starting and ending positions according to the time series. Then, the local neighborhood analysis window moves point-by-point from the first starting position to the first ending position, reading the sensor data point values ​​and sorting them in descending order to obtain a numerical sequence. Next, the midpoint of the numerical sequence is located, and the first value is extracted as the local central tendency representation value of the local neighborhood analysis window. Simultaneously, a discrete analysis is performed on the sensor data point value set through the local neighborhood analysis window to calculate the numerical dispersion. Finally, the local central tendency representation value is combined with the numerical dispersion for local analysis to obtain the local data distribution characteristics.

[0023] Next, instantaneous anomaly fluctuation analysis is performed based on local data distribution characteristics to extract instantaneous anomaly jump points. In this process, each sensor data point in the time-series aligned dataset is traversed, and the local data distribution characteristics of each sensor data point's corresponding location are read. Then, the value of the current sensor data point is compared with the local central trend characterization value, and the numerical dispersion is used to determine whether the current sensor data point is located within the corresponding local data distribution interval. When the value of the current sensor data point is not located within the local data distribution interval centered on the local central trend characterization value and characterized by numerical dispersion, the location corresponding to the current sensor data point is determined as an instantaneous anomaly jump point. An instantaneous anomaly jump point refers to a data location in the time series that does not conform to the corresponding local data distribution state. Through the above processing, the locations corresponding to abnormal data are extracted from the time-series aligned dataset.

[0024] Next, the time-series aligned dataset is cleaned based on the transient abnormal jump points. The data positions corresponding to each transient abnormal jump point are located in the time-series aligned dataset, and each data position is used as anomaly handling position for data cleaning. Here, data cleaning refers to removing data that does not conform to the corresponding local data distribution state from the valid analysis data or marking it as a missing position. During processing, data that has not been identified as transient abnormal jump points are retained, and data that has been identified as transient abnormal jump points are removed or marked as missing, so that the abnormal data in the time-series aligned dataset is isolated, thereby obtaining a cleaned time-series dataset.

[0025] Finally, the time-series dataset is traversed to identify gaps and determine multiple data gap locations. Based on the time-series dataset, an adjacency search is performed according to the multiple data gap locations to extract multiple valid sensor data points. These valid sensor data points are then filled into the multiple data gap locations to generate a fused time-series data stream. Specifically, firstly, based on the cleaned time-series dataset, second start position information and second end position information are defined according to the time sequence. Then, through a local neighborhood analysis window, the data storage locations of the cleaned time-series dataset are sequentially traversed from the second start position information to the second end position information to determine whether there are empty placeholder identifiers. When empty placeholder identifiers exist, multiple data gap locations to be filled are extracted, and the data is searched forward and backward according to the time sequence to obtain the first effective sensing data point and the second effective sensing data point. Subsequently, the first distance value between the first effective sensing data point and the target data gap location to be filled, and the second distance value between the second effective sensing data point and the target data gap location to be filled are determined respectively. Finally, based on the comparison result of the first distance value and the second distance value, the corresponding effective sensing data point is selected to fill in the data gap location to be filled, thereby generating a fused time-series data stream.

[0026] Furthermore, in the method provided in the application embodiments, traversing the time-aligned dataset to perform local data analysis and obtain local data distribution characteristics further includes: Based on the time-series aligned dataset, a local sliding analysis is performed according to the time series to define a local neighborhood analysis window. First starting position information and first ending position information are defined according to the time series based on the time series aligned dataset. The local neighborhood analysis window moves point-by-point from the first starting position information to the first ending position information, reading the set of sensor data point values. The set of sensor data point values ​​is sorted in descending order of value size to obtain a value sequence. The midpoint of the value sequence is located, and a first value is extracted as the local central trend representation value of the local neighborhood analysis window. Discrete analysis is performed on the set of sensor data point values ​​through the local neighborhood analysis window to calculate the numerical dispersion. The local central trend representation value is combined with the numerical dispersion for local analysis to obtain the local data distribution characteristics.

[0027] In this embodiment of the application, local sliding analysis is performed based on the time-series aligned dataset. When defining the local neighborhood analysis window, the time-series aligned dataset is written into the sequential storage array in chronological order, and the window length of the local neighborhood analysis window is preset so that the local neighborhood analysis window covers a fixed number of continuous sensor data points in each processing. Then, starting from the current time position of the time-series aligned dataset, a continuous data segment with the same window length is extracted as the local neighborhood analysis window currently participating in the analysis.

[0028] Next, based on the time-series aligned dataset, the first start position information and the first end position information are defined according to the time series. Specifically, the position where the local neighborhood analysis window first completely falls into the time-series aligned dataset is determined as the first start position information, and the position where the local neighborhood analysis window last completely falls into the time-series aligned dataset is determined as the first end position information. In actual processing, the first valid data position in the time-series aligned dataset is used as the first start position information, and the (N-W+1)th valid data position is used as the first end position information, where N represents the total number of data in the time-series aligned dataset, and W represents the window length of the local neighborhood analysis window.

[0029] Then, the local neighborhood analysis window moves point by point from the first starting position information to the first ending position information, reading the set of sensor data point values. Specifically, the local neighborhood analysis window starts from the position corresponding to the first starting position information and moves one data position backward along the time series direction each time. After each movement, all sensor data point values ​​within the current coverage area of ​​the local neighborhood analysis window are read sequentially, and all read sensor data point values ​​are written to a cache array in turn to form the set of sensor data point values ​​corresponding to the current window position. Then, the window movement and data reading operations are repeated until the position corresponding to the first ending position information is reached.

[0030] The sensor data point values ​​are then sorted in descending order of magnitude. The sensor data point values ​​corresponding to the current window position are then sorted by comparing the magnitudes of all sensor data point values ​​and rearranging them in descending order. After sorting, the rearranged results are written into the result array to obtain the numerical sequence corresponding to the current local neighborhood analysis window.

[0031] Next, the midpoint of the numerical sequence is located, and the first value is extracted as the local central trend representation value of the local neighborhood analysis window. Specifically, after obtaining the numerical sequence, the length L of the numerical sequence is read, and the midpoint position is calculated according to the position index; when L is odd, the value corresponding to the (L+1) / 2th position is directly taken as the first value; when L is even, the value corresponding to the L / 2th position is taken as the first value; then the first value is written into the position record corresponding to the current local neighborhood analysis window as the local central trend representation value of that window position.

[0032] Discrete analysis is performed on the set of sensor data points through a local neighborhood analysis window. When calculating the numerical dispersion, the local central tendency characterization value corresponding to the current local neighborhood analysis window is used as the benchmark value. The difference between the value of each sensor data point in the set of sensor data points and the local central tendency characterization value is calculated in turn. The absolute values ​​of each difference are taken and accumulated. Then, the accumulated result is divided by the number of data points in the set of sensor data points to obtain the numerical dispersion corresponding to the current local neighborhood analysis window.

[0033] Finally, the local central tendency value is combined with the numerical dispersion for local analysis. Specifically, at the position corresponding to the current local neighborhood analysis window, the local central tendency value and the numerical dispersion are read simultaneously. The local central tendency value is written as the preceding data and the numerical dispersion is written as the following data into the same feature record to form the local data distribution feature corresponding to the current window position. Then, this writing process is repeated along all window positions from the first starting position information to the first ending position information to obtain the local data distribution features of the time-series aligned dataset at each time position.

[0034] Furthermore, in the method provided in the application embodiment, the process of traversing the cleaned time-series dataset to identify gaps and determine multiple data gap locations, performing an adjacency search based on the cleaned time-series dataset according to the multiple data gap locations to extract multiple valid sensor data points, filling the multiple valid sensor data points into the multiple data gap locations, and generating the fused time-series data stream further includes: Based on the cleaned time-series dataset, second start position information and second end position information are defined according to the time sequence. The local neighborhood analysis window sequentially traverses from the second start position information to the second end position information to determine if there are any empty space placeholders in the data storage locations of the cleaned time-series dataset. When an empty space placeholder exists in the data storage location of the cleaned time-series dataset, a record extraction instruction is generated to extract multiple data gaps to be filled. The multiple data gaps to be filled are then searched forward according to the time sequence to obtain a first valid sensor data point, which includes a first distance value between the first valid sensor data point and the target data gap to be filled. The system searches backward in time sequence according to the multiple data gaps to be filled, obtaining second valid sensor data points. Each second valid sensor data point contains a second distance value between itself and the target data gap. When the first distance value is greater than the second distance value, the second valid sensor data point is selected for data completion to generate a fused time-series data stream. When the first distance value is less than the second distance value, the first valid sensor data point is selected for data completion to generate a fused time-series data stream. When the first distance value is equal to the second distance value, either the first valid sensor data point or the second valid sensor data point is randomly selected for data completion to generate the fused time-series data stream.

[0035] In this embodiment, the time-series data of cleaning is written into a continuous storage space according to the time sequence, and a continuous number is configured for each data storage location according to the time sequence; the starting number in the continuous storage space is read and determined as the second starting position information, and the ending number in the continuous storage space is read and determined as the second ending position information, thereby limiting the start and end range of the gap detection along the time sequence.

[0036] The local neighborhood analysis window is positioned at the data storage location corresponding to the second starting position information, and moves sequentially along the time series according to the numbering to the data storage location corresponding to the second ending position information. Each time the local neighborhood analysis window moves to a data storage location, the stored content in that data storage location is read, and the reading result is matched with the preset empty space placeholder identifier to determine whether there is an empty space placeholder identifier in that data storage location.

[0037] When a placeholder identifier exists at any data storage location, a record extraction instruction is generated, and the location number, time location, and vacancy mark corresponding to that data storage location are written into the vacancy record table. The instruction is executed sequentially until the data storage location corresponding to the second end location information is reached, thereby extracting all data vacancy locations to be filled from the vacancy record table.

[0038] For each data gap to be filled, one of the gaps is selected as the target gap and a reverse search is performed forward along the time series. During the search, the data storage locations before the target gap are read one by one, and the placeholder identifier is continuously checked until the first data storage location without a placeholder identifier is located. The sensor data corresponding to this data storage location is determined as the first valid sensor data point. At the same time, the first distance value is obtained by subtracting the location number of the first valid sensor data point from the location number of the target gap.

[0039] Around the same target data gap to be filled, a sequential search is performed forward along the time series. During the search, the data storage locations after the target data gap to be filled are read one by one, and the empty space placeholder identifier is continuously judged until the first data storage location without an empty space placeholder identifier is located. The sensor data corresponding to the data storage location is determined as the second valid sensor data point. At the same time, the location number of the target data gap to be filled is subtracted from the location number of the second valid sensor data point to obtain the second distance value.

[0040] When the first distance value is greater than the second distance value, after comparing and judging the first distance value and the second distance value, the second valid sensor data point is selected as the data source for filling, and the data value of the second valid sensor data point is written into the target data gap position to be filled, overwriting the original empty space placeholder identifier at that position. In this way, the data filling of the target data gap position to be filled is completed, and the corresponding data position in the fused time-series data stream is formed.

[0041] When the first distance value is less than the second distance value, after comparing and judging the first distance value and the second distance value, the first valid sensor data point is selected as the data source for filling, and the data value of the first valid sensor data point is written into the target data gap position to be filled, overwriting the original empty space placeholder identifier at that position, thereby completing the data filling of the target data gap position to be filled and forming the corresponding data position in the fused time-series data stream.

[0042] When the first distance value equals the second distance value, a random selection process is invoked to generate a selection result between the first and second valid sensor data points. When the selection result corresponds to the first valid sensor data point, the data value of the first valid sensor data point is written to the target data gap to be filled. When the selection result corresponds to the second valid sensor data point, the data value of the second valid sensor data point is written to the target data gap to be filled. This process is repeated to complete the data filling of all data gaps to be filled, generating a fused time-series data stream.

[0043] Step S300: Construct a dual-layer feature extraction window. The dual-layer feature extraction window mechanism includes a decay window and a dynamic sliding window. Real-time feature fusion extraction is performed on the fused time-series data stream through the decay window and the dynamic sliding window to obtain multi-scale real-time feature vectors.

[0044] In this embodiment, when constructing a dual-layer feature extraction window, a window mechanism for real-time feature fusion extraction is first established for the fused time-series data stream. This window mechanism is divided into a decay window and a dynamic sliding window. The decay window is used to perform weighted feature calculation by combining historical time-series data segments within the historical data observation interval, and the dynamic sliding window is used to perform window size adjustment and real-time feature extraction by combining the real-time changes of the fused time-series data stream, thereby forming a dual-layer feature extraction window.

[0045] After the dual-layer feature extraction window is constructed, the fused time-series data stream is weighted and calculated using a decay window combined with a decay weight coefficient to generate a first-scale real-time feature vector. Simultaneously, a dynamic sliding window is used to perform real-time feature fusion extraction on the fused time-series data stream according to the window size parameters to obtain a second-scale real-time feature vector. The first-scale real-time feature vector and the second-scale real-time feature vector are then used as multi-scale real-time feature vectors, thus completing the process of real-time feature fusion extraction of the fused time-series data stream through a dual-layer feature extraction window.

[0046] Furthermore, in the method provided in the application embodiments, the real-time feature fusion extraction of the fused time-series data stream through the attenuation window to obtain a multi-scale real-time feature vector further includes: The memory buffer of the intelligent fusion terminal is accessed and defined to determine the historical data observation interval; historical time-series data segments are extracted based on the historical data observation interval; spatial capacity extreme value analysis is performed based on the memory buffer, and window capacity parameters are set according to the capacity extreme values; data attenuation weights are allocated to the attenuation window based on the historical time-series data segments to determine the attenuation weight coefficient; weighted feature calculation is performed on the fused time-series data stream through the attenuation window and the attenuation weight coefficient to generate a first-scale real-time feature vector.

[0047] In this embodiment, the memory buffer of the intelligent fusion terminal is first retrieved and delineated to determine the historical data observation interval. Specifically, the starting address, current write address, and available storage length of the memory buffer are read first, and then the memory buffer is delineated according to a preset historical retention range. When the preset historical retention range is a time length, the time corresponding to the current write address is taken as the end point of the interval, and the starting address corresponding to the preset time length is determined forward along the time sequence. When the preset historical retention range is a number of data entries, the data position corresponding to the current write address is taken as the end point of the interval, and the starting address corresponding to the preset number of data entries is determined forward along the time sequence. Subsequently, the continuous address range between the starting address and the end point of the interval is determined as the historical data observation interval.

[0048] Next, historical time-series data segments are extracted based on historical data observation intervals. Data is read sequentially according to the start and end addresses corresponding to the historical data observation intervals, retrieving fused time-series data entries one by one from the memory buffer within each interval, while simultaneously reading the timestamps corresponding to each fused time-series data entry. The read results are then written into a historical data cache sequence in chronological order, ensuring that the first data entry in the historical data cache sequence corresponds to the earliest time and the last data entry corresponds to the historical time closest to the current write position. After completing the reading of data across the entire address range, this historical data cache sequence is determined as a historical time-series data segment.

[0049] Subsequently, a space capacity limit analysis is performed based on the memory buffer, and the window capacity parameter is set according to the capacity limit. Specifically, the available storage space allocated to the decay window is first counted, then the storage length corresponding to a single fused time series data is read, and the maximum number of data rows that the decay window can accommodate under the current memory conditions is calculated based on the correspondence between the available storage space and the storage length of a single fused time series data. This maximum number of data rows is then determined as the capacity limit, and the capacity limit is compared with the actual number of data rows in the historical time series data segment. When the capacity limit is not less than the actual number of data rows, the actual number of data rows is determined as the window capacity parameter; when the capacity limit is less than the actual number of data rows, the capacity limit is determined as the window capacity parameter.

[0050] The attenuation weights of the attenuation window are assigned based on historical time-series data segments to determine the attenuation weight coefficients. Specifically, firstly, a number of consecutive historical data points corresponding to the window capacity parameter are extracted from the historical time-series data segments and arranged in chronological order to form the data input sequence for the attenuation window. Then, weights are assigned according to the time proximity of the data in the data input sequence. The historical data point at the end of the data input sequence, closest to the current time, is assigned the largest original weight, the historical data point preceding it is assigned the second largest original weight, and the remaining historical data points are assigned progressively decreasing original weights along the time sequence until the historical data point at the beginning of the data input sequence, farthest from the current time, is assigned the smallest original weight. After all the original weights are determined, the original weight corresponding to each historical data point is divided by the sum of all the original weights to obtain the attenuation weight coefficient for each historical data point. These attenuation weight coefficients are then written into the weight cache sequence in chronological order.

[0051] Finally, weighted feature calculation is performed on the fused time-series data stream using an attenuation window combined with attenuation weight coefficients. In this process, the fused time-series data entering the attenuation window at the current time is first read, and this fused time-series data is combined with historical time-series data segments within the attenuation window to form a calculation sequence in chronological order. Then, the value of each data point in the calculation sequence is multiplied by its corresponding attenuation weight coefficient, and all multiplication results are accumulated along the time sequence to obtain the weighted feature result corresponding to the attenuation window. When the fused time-series data stream contains more than one feature dimension, the same corresponding multiplication and accumulation processes are performed on each feature dimension to obtain the weighted feature result corresponding to each feature dimension. Finally, the weighted feature results are written into the feature vector storage area according to a preset dimension order to form the first-scale real-time feature vector.

[0052] Furthermore, in the method provided in the application embodiments, the real-time feature fusion extraction of the fused time-series data stream through the dynamic sliding window to obtain multi-scale real-time feature vectors further includes: Based on the memory buffer, memory is allocated according to the dynamic sliding window to generate an independent memory address space; data containment analysis is performed based on the independent memory address space, and an initial window size parameter is set; based on the initial window size parameter, adjacent data in the fused time-series data stream are continuously read, and the numerical difference is calculated; the time sampling interval data of adjacent data is extracted, and the numerical difference is divided by the time sampling interval data to calculate the real-time change rate; based on the real-time change rate, data stream change analysis is performed to generate a data stream change trend, and the initial window size parameter of the dynamic sliding window is dynamically adjusted to generate a window size parameter; through the dynamic sliding window according to the window size parameter, real-time feature fusion extraction is performed on the fused time-series data stream to obtain a second-scale real-time feature vector.

[0053] In this embodiment of the application, when allocating memory based on the memory buffer according to the dynamic sliding window, the available contiguous storage area in the memory buffer is first read, and according to the data length that the dynamic sliding window needs to accommodate in one processing, a contiguous and unoccupied storage area is extracted from the memory buffer as the data storage range of the dynamic sliding window; then the start address and end address of the storage area are determined as the independent memory address space corresponding to the dynamic sliding window, so that the fused timing data subsequently read by the dynamic sliding window is continuously written into the independent memory address space.

[0054] Next, data storage analysis is performed based on the independent memory address space, and an initial window size parameter is set. Specifically, the total storage length of the independent memory address space is first calculated, then the storage length corresponding to a single fused time-series data is read, and based on the correspondence between the total storage length of the independent memory address space and the storage length of a single fused time-series data, the number of data entries that the independent memory address space can continuously accommodate under the current conditions is calculated. Then, this number of data entries is compared with the preset initial window length. When the number of data entries is not less than the preset initial window length, the preset initial window length is determined as the initial window size parameter; when the number of data entries is less than the preset initial window length, the number of data entries is determined as the initial window size parameter. This determines the data coverage range when the dynamic sliding window begins to perform real-time feature fusion extraction.

[0055] Then, based on the initial window size parameters, adjacent data in the fused time-series data stream are continuously read, and the numerical difference is calculated. In this process, continuous data is sequentially read from the fused time-series data stream according to the initial window size parameters and written into independent memory address spaces in chronological order. Within the data range covered by the current dynamic sliding window, adjacent pairs of fused time-series data form a comparison unit. The values ​​of the preceding and following fused time-series data are extracted, and the value of the preceding fused time-series data is subtracted from the value of the following fused time-series data to obtain the numerical difference corresponding to this group of adjacent data. This difference calculation is continuously performed along all adjacent data covered by the current dynamic sliding window to obtain a sequence of numerical differences.

[0056] Next, the time sampling interval data of adjacent data is extracted, and the numerical difference is divided by the time sampling interval data to calculate the real-time change rate. Specifically, for each group of adjacent data with obtained numerical differences, the sampling time corresponding to the previous fused time series data and the sampling time corresponding to the next fused time series data in that group of adjacent data are read, and the previous sampling time is subtracted from the next sampling time to obtain the time sampling interval data corresponding to that group of adjacent data; then, the numerical difference corresponding to that group of adjacent data is divided by the corresponding time sampling interval data to obtain the real-time change rate corresponding to that group of adjacent data; this process is repeated for all adjacent data within the coverage area of ​​the current dynamic sliding window to obtain the real-time change rate sequence.

[0057] Subsequently, data stream change analysis is performed based on the real-time change rate to generate a data stream change trend. The initial window size parameter of the dynamic sliding window is dynamically adjusted to generate the window size parameter. In this process, all rate values ​​in the real-time change rate sequence are first statistically analyzed to obtain the average rate value and the maximum rate value within the coverage area of ​​the current dynamic sliding window. Then, the average rate value and the maximum rate value are compared with the preset rate upper limit and preset rate lower limit, respectively. When both the average rate value and the maximum rate value are greater than the preset rate upper limit, the fused time-series data stream is determined to be in a rapid change state, and the initial window size parameter is reduced by one data position. When both the average rate value and the maximum rate value are less than the preset rate lower limit, the fused time-series data stream is determined to be in a slow change state, and the initial window size parameter is increased by one data position. When the average rate value and the maximum rate value are between the preset rate upper limit and the preset rate lower limit, the fused time-series data stream is determined to be in a stable change state, and the initial window size parameter remains unchanged. The adjusted result is then determined as the window size parameter. To prevent the adjusted window size parameter from exceeding the capacity of the independent memory address space, if the window size parameter is less than the preset minimum window length after reducing one data location, the preset minimum window length will be used as the window size parameter. If the window size parameter is greater than the number of data rows that the independent memory address space can accommodate after increasing one data location, the number of data rows that can be accommodated will be used as the window size parameter.

[0058] Finally, a dynamic sliding window is used to perform real-time feature fusion and extraction on the fused time-series data stream according to the window size parameters. In this process, a continuous data segment of corresponding length is extracted from the fused time-series data stream according to the window size parameters as the analysis range of the current dynamic sliding window, and it slides position by position along the time series as the fused time-series data stream continues to input. After each slide, the fused time-series data, numerical differences, time sampling interval data, and real-time change rate within the current dynamic sliding window's coverage area are read, and features are calculated according to a preset feature dimension order. The preset feature dimension order can be set sequentially as: fused time-series data feature dimension, numerical differences feature dimension, time sampling interval feature dimension, and real-time change rate feature dimension. Under the fused time-series data feature dimension, all fused time-series data within the current dynamic sliding window's coverage area are summed and divided by the number of data points covered by the current dynamic sliding window to obtain the average value of the fused time-series data. Under the numerical differences feature dimension, the current dynamic... The average numerical difference is obtained by summing all numerical differences within the coverage area of ​​the sliding window and dividing by the number of adjacent data pairs covered by the current dynamic sliding window. Under the time sampling interval feature dimension, the average time sampling interval data within the coverage area of ​​the current dynamic sliding window is summed and divided by the number of adjacent data pairs covered by the current dynamic sliding window. Under the real-time change rate feature dimension, the average real-time change rate is summed and divided by the number of adjacent data pairs covered by the current dynamic sliding window. Then, according to the preset feature dimension order, the average value of the fused time-series data, the average value of the numerical difference, the average value of the time sampling interval, and the average value of the real-time change rate are sequentially written into the feature vector storage area to form the second-scale real-time feature vector corresponding to the current dynamic sliding window.

[0059] Step S400: Construct an initial feature vector matrix based on the multi-scale real-time feature vectors, perform dynamic feature weight allocation and update, and obtain the feature vector matrix.

[0060] In this embodiment of the application, when constructing an initial feature vector matrix based on multi-scale real-time feature vectors for dynamic feature weight allocation and update, firstly, dimensional analysis is performed based on multi-scale real-time feature vectors to extract multiple feature dimensions; then, the intelligent fusion terminal retrieves the number of historical features retained, and uses the number of historical features retained as the number of rows and the multiple feature dimensions as the number of columns to construct a two-dimensional array space; subsequently, the two-dimensional array space is filled with data to construct the initial feature vector matrix.

[0061] After the initial feature vector matrix is ​​constructed, a one-dimensional weight storage array is created based on multiple feature dimensions. The initial feature vector matrix is ​​then weighted using the one-dimensional weight storage array to determine multiple initial weight arrays. The contribution of the initial feature vector matrix is ​​then calculated by traversing the initial feature vector matrix to obtain multiple contribution scores. The multiple initial weight arrays are then iteratively updated based on the multiple contribution scores to generate multiple weight arrays. Finally, the initial feature vector matrix is ​​reweighted according to the multiple weight arrays to generate the feature vector matrix.

[0062] Furthermore, in the method provided in the application embodiments, the process of constructing an initial feature vector matrix based on the multi-scale real-time feature vectors and performing dynamic feature weight allocation and updating to obtain the feature vector matrix further includes: Dimensional analysis is performed based on the multi-scale real-time feature vectors to extract multiple feature dimensions. The number of historical features retained is retrieved from the intelligent fusion terminal, and this number is used as the number of rows, while the multiple feature dimensions are used as the number of columns to construct a two-dimensional array space. Data is filled into the two-dimensional array space to construct an initial feature vector matrix. A one-dimensional weight storage array is created based on the multiple feature dimensions, and weights are assigned to the initial feature vector matrix using this one-dimensional weight storage array to determine multiple initial weight arrays. The initial feature vector matrix is ​​traversed to calculate contribution, obtaining multiple contribution scores. The multiple initial weight arrays are iteratively updated according to the multiple contribution scores to generate multiple weight arrays. The initial feature vector matrix is ​​then reweighted according to the multiple weight arrays to generate the feature vector matrix.

[0063] In this embodiment of the application, when performing dimensional analysis based on multi-scale real-time feature vectors, the multi-scale real-time feature vectors generated at the current moment are first read, and sequential parsing is performed according to the arrangement order of the feature values ​​in the multi-scale real-time feature vectors. During the parsing process, each feature component contained in the multi-scale real-time feature vector is identified item by item, and the total number of feature components is counted, which is determined as the number of feature dimensions. Subsequently, according to the arrangement position in the multi-scale real-time feature vectors, each feature component is assigned a corresponding dimension position in sequence, thereby completing the dimensional analysis and extracting multiple feature dimensions.

[0064] Next, based on the intelligent fusion terminal, the number of historical features retained is retrieved. This number is used as the number of rows, and multiple feature dimensions are used as the number of columns to construct a two-dimensional array space. Specifically, the intelligent fusion terminal first reads the number of historical features retained from the preset storage parameters and determines this number as the number of rows in the two-dimensional array space, which limits the number of historical feature vectors that can be continuously retained. Then, the number of multiple feature dimensions extracted in the previous step is determined as the number of columns in the two-dimensional array space, so that each column in the two-dimensional array space corresponds to one feature dimension. Subsequently, corresponding contiguous storage areas are allocated in the memory according to the number of rows and columns, and the two-dimensional data positional relationship is established in a row-column manner to form a two-dimensional array space. In this space, each row is used to store a multi-scale real-time feature vector corresponding to a certain moment, and each column is used to store the feature values ​​corresponding to the same feature dimension at different moments, thus completing the construction of the two-dimensional array space.

[0065] Next, data is filled into the two-dimensional array space to construct the initial feature vector matrix. During this process, the two-dimensional array space is traversed, and it is determined whether the number of rows filled with feature vectors has reached the historical feature retention limit. When the number of rows filled with feature vectors has not reached the historical feature retention limit, multi-scale real-time feature vectors are filled into the first empty row of the two-dimensional array space according to multiple feature dimensions. When the number of rows filled with feature vectors reaches the historical feature retention limit, the feature vector data of the row with the longest storage time in the two-dimensional array space is removed entirely, and the remaining rows are shifted forward one row to fill the empty spaces. Then, multi-scale real-time feature vectors are filled into the resulting empty rows. By continuously executing the above data filling process, the initial feature vector matrix is ​​finally constructed.

[0066] A one-dimensional weight storage array is created based on multiple feature dimensions. Weights are then assigned to the initial feature vector matrix using this array. To determine multiple initial weight arrays, the column number of the initial feature vector matrix is ​​first read and used as the number of feature dimensions. A one-dimensional weight storage array is then created according to the number of feature dimensions, ensuring that each array position in the one-dimensional weight storage array corresponds one-to-one with a column of feature dimensions in the initial feature vector matrix. Initial weight values ​​are then assigned to each array position in the one-dimensional weight storage array. When all feature dimensions participate in weight allocation, the total weight is evenly distributed across the multiple feature dimensions, ensuring that the initial weight values ​​corresponding to each array position are the same, with a total weight of one. After the one-dimensional weight storage array is created, the feature vector data of each row in the initial feature vector matrix is ​​read row by row, and the initial weight values ​​in the one-dimensional weight storage array are copied and distributed to each feature dimension corresponding to each row in the column direction. This ensures that each row in the initial feature vector matrix forms a set of initial weight arrays corresponding to multiple feature dimensions.

[0067] The contribution is then calculated by traversing the initial feature vector matrix. Specifically, the matrix is ​​traversed row by row, and feature values ​​are read column by column within the current row, along with the weight values ​​in the corresponding initial weight array. The feature value for each feature dimension in the current row is then multiplied by its corresponding initial weight value to obtain the weighted result for that feature dimension. The absolute values ​​of all weighted results for all feature dimensions in the current row are then summed to obtain the total weighted value for the current row. Finally, the absolute value of the weighted result for a specific feature dimension in the current row is divided by the total weighted value for the current row to obtain the contribution score for that feature dimension in the current row. This process is repeated for all feature dimensions in the current row, and then repeated for all rows in the initial feature vector matrix to obtain multiple contribution scores.

[0068] Next, multiple initial weight arrays are iteratively updated based on multiple contribution scores. In this process, the contribution score for each row is first read row by row, and all contribution scores in that row are normalized so that the sum of all normalized contribution scores in that row is one. Then, each initial weight value in the initial weight array of that row is matched column by column with the corresponding normalized contribution score, and an update calculation is performed for each feature dimension. The initial weight value corresponding to that feature dimension is added to the normalized contribution score corresponding to that feature dimension, and the sum is divided by two to obtain the updated weight value corresponding to that feature dimension. After all feature dimensions in the current row have obtained updated weight values, all updated weight values ​​in that row are normalized again so that the sum of all updated weight values ​​in that row is one, thus obtaining the weight array corresponding to that row. Subsequently, the above update process is repeated for all rows in the initial feature vector matrix, and this update process is repeated continuously for a preset number of iterations, thereby generating multiple weight arrays.

[0069] Finally, the initial eigenvector matrix is ​​reweighted according to multiple weight arrays to generate the eigenvector matrix. In this process, row-by-row eigenvector data is read sequentially along the row direction of the initial eigenvector matrix, and the corresponding weight array is read simultaneously. Then, for each feature dimension in the current row, the eigenvalue corresponding to that feature dimension is multiplied by the updated weight value corresponding to that feature dimension to obtain the reweighted eigenvalue. After the product calculation is completed for all feature dimensions in the current row, each reweighted eigenvalue is written back to its corresponding matrix position in the original column order. This process is repeated for all rows in the initial eigenvector matrix to obtain the reweighted eigenvector matrix, which is then designated as the eigenvector matrix.

[0070] Furthermore, in the method provided in the application embodiments, filling the two-dimensional array space with data to construct an initial feature vector matrix further includes: The data filling judgment is performed by traversing the two-dimensional array space: S1: Determine whether the number of rows filled with feature vectors in the two-dimensional array space has reached the number of historical features to be retained; S2: When the number of rows filled with feature vectors in the two-dimensional array space has not reached the number of historical features to be retained, the multi-scale real-time feature vectors are filled into the first empty row of the two-dimensional array space according to the multiple feature dimensions; S3: When the number of rows filled with feature vectors in the two-dimensional array space has reached the number of historical features to be retained, the row feature vector data with the longest storage time maximum value in the two-dimensional array space is removed as a whole, and the remaining rows are shifted forward one row to fill the empty space, obtaining empty row spaces; S4: The multi-scale real-time feature vectors are filled into the empty row spaces of the two-dimensional array space; Based on the above steps S1, S2, S3, and S4, continuous filling is performed to construct the initial feature vector matrix.

[0071] In this embodiment of the application, when traversing the two-dimensional array space to determine data filling, the writing status of each row is read sequentially from front to back along the row direction of the two-dimensional array space, and the rows that have been written with complete multi-scale real-time feature vectors are counted. After the status reading of all rows is completed, the number of rows with filled feature vectors is compared with the number of historical features retained to determine whether the number of rows with filled feature vectors in the two-dimensional array space has reached the number of historical features retained.

[0072] When the number of rows in the two-dimensional array space that have been filled with feature vectors has not reached the number of historical features to be retained, the system continues to sequentially search for empty rows that have not yet been written with feature values ​​along the row direction of the two-dimensional array space, and determines the first empty row as the first empty row. Then, the system reads the multi-scale real-time feature vector to be written, and writes each feature value in the multi-scale real-time feature vector into the corresponding column storage location of the first empty row according to the arrangement order of multiple feature dimensions in the column direction. This allows the feature values ​​in the row to be filled sequentially from the first column to the last column, thereby filling the first empty row of the two-dimensional array space with the multi-scale real-time feature vector.

[0073] When the number of rows filled with feature vectors in the two-dimensional array space reaches the number of historical features to be retained, the storage time corresponding to each filled row is read first, and the storage times of each row are compared to locate the row containing the feature vector data with the longest storage time. Then, the entire row is cleared, and all feature values ​​in each column of the row are removed from the storage location. After the entire row is removed, the remaining rows following it are moved forward one by one in ascending order of row number, so that the data of the current row is copied to the position corresponding to the previous row, and the original position of the current row is cleared, until the last row is moved forward, thus creating a space row empty at the end of the two-dimensional array space.

[0074] After obtaining the empty space in the space, the current multi-scale real-time feature vector to be written is read, and according to the arrangement order of multiple feature dimensions in the column direction, each feature value in the multi-scale real-time feature vector is written to the corresponding column storage position of the empty space in the space. During the writing process, the first feature value in the multi-scale real-time feature vector is written to the first column of the empty space, and the subsequent feature values ​​are written to the corresponding subsequent columns in sequence, until the last feature value is written to the last column, thus completing the filling of the multi-scale real-time feature vector into the empty space of the two-dimensional array space.

[0075] As new multi-scale real-time feature vectors are continuously generated, the following processing steps are repeatedly executed: determining the number of rows filled with feature vectors, writing the first empty row, removing the feature vector data of the row with the longest storage time, shifting the remaining rows forward one row to fill empty spaces, and writing empty spaces in the space rows. This ensures that each row in the two-dimensional array space continuously retains the multi-scale real-time feature vectors generated at the most recent time, and that each column continuously corresponds to the feature values ​​of the same feature dimension at different times. During the continuous filling process, the construction of the two-dimensional array space into the initial feature vector matrix is ​​completed.

[0076] Step S500: Based on the feature vector matrix, perform state recognition response on the intelligent fusion terminal and generate auxiliary decision-making instructions for the intelligent fusion terminal.

[0077] In this embodiment, when the intelligent fusion terminal performs state recognition response based on the feature vector matrix, the feature vector matrix is ​​first synchronized to the state classifier built into the intelligent fusion terminal for analysis to generate multiple state categories; then, confidence evaluation is performed based on the multiple state categories to generate multiple confidence score values, and the target confidence score value is determined by extracting the maximum value of the multiple confidence score values; subsequently, the target confidence score value is matched with the multiple state categories to determine the target state category, and the target state category is used as candidate state data; finally, hierarchical decision-making is performed based on the candidate state data to obtain multi-level auxiliary decision messages, and the multi-level auxiliary decision messages are associated and encapsulated to construct auxiliary decision instructions.

[0078] Furthermore, in the method provided in the application embodiments, the process of performing state recognition response on the intelligent fusion terminal based on the feature vector matrix and generating auxiliary decision-making instructions for the intelligent fusion terminal further includes: The feature vector matrix is ​​synchronized to the state classifier built into the intelligent fusion terminal for analysis to generate multiple state categories; confidence assessment is performed based on the multiple state categories to generate multiple confidence scores; the maximum value is extracted from the multiple confidence scores to determine the target confidence score; the target confidence score is matched with the multiple state categories to determine the target state category, and the target state category table is used as candidate state data; hierarchical decision-making is performed using the candidate state data to obtain multi-level auxiliary decision messages; the multi-level auxiliary decision messages are associated and encapsulated to construct auxiliary decision instructions.

[0079] In this embodiment, the state classifier is trained before being written to the intelligent fusion terminal. Specifically, historical feature vector matrices under multiple known states are first collected, and each historical feature vector matrix is ​​labeled with its corresponding real state category to form training samples. Then, each historical feature vector matrix is ​​expanded into a one-dimensional input sequence in row-major order. The state classifier adopts a multi-class linear classification structure. The input data is a one-dimensional input sequence, and the number of state categories is a preset number. Each state category corresponds to a set of classification weights and a bias value. At the start of training, the classification weights corresponding to each state category are initialized to zero or a small unified initial value, and the bias value is initialized to zero. Subsequently, for each training sample, the classification score corresponding to each state category is calculated. The calculation method is to multiply each input value in the one-dimensional input sequence with each classification weight corresponding to a certain state category, then sum all the product results, and finally add the bias value corresponding to the state category to obtain the classification score corresponding to the state category. After completing the above calculation for all state categories, the state category with the highest score is selected as the current predicted state category. If the current predicted state category matches the true state category, the classification weights and biases corresponding to that training sample remain unchanged. If the current predicted state category does not match the true state category, the classification weights corresponding to the true state category are progressively increased by a correction amount, and the classification weights corresponding to the current predicted state category are progressively decreased by a correction amount. Each correction amount is obtained by multiplying the corresponding input value by a preset learning rate. At the same time, the biases corresponding to the true state category are increased by a preset learning rate, and the biases corresponding to the current predicted state category are decreased by a preset learning rate. The above input, score calculation, state determination, and parameter correction process is repeated for all training samples until a preset number of training rounds are completed or the number of classification errors is reduced to a preset range, thereby obtaining the trained state classifier.

[0080] When synchronizing the feature vector matrix to the built-in state classifier of the intelligent fusion terminal for analysis, all matrix elements in the feature vector matrix are first read, and each matrix element is expanded into a one-dimensional input sequence in row-major order. This one-dimensional input sequence is then input into the trained state classifier. In the state classifier, each state category corresponds to a set of trained classification weights and a bias value. Subsequently, a classification score is calculated for each state category. The calculation method is as follows: the first input value in the one-dimensional input sequence is multiplied by the first classification weight corresponding to that state category; the second input value is multiplied by the second classification weight corresponding to that state category; and so on, until all input values ​​are multiplied by their corresponding classification weights. All product results are then accumulated, and the bias value corresponding to that state category is added to the accumulated result to obtain the classification score for that state category. After repeating the above calculation for all state categories, multiple state categories corresponding one-to-one with each classification score are output.

[0081] Next, when performing confidence assessment based on multiple state categories, the classification scores corresponding to all state categories are first read, and the minimum classification score is determined from all classification scores. When the minimum classification score is less than zero, the absolute value of the minimum classification score is incremented by one as a shift amount, and this shift amount is simultaneously added to all classification scores to obtain the adjusted classification scores corresponding to all state categories. When the minimum classification score is greater than or equal to zero, one is simultaneously added to all classification scores to obtain the adjusted classification scores corresponding to all state categories. Then, all adjusted classification scores are accumulated to obtain the total score value. Next, the adjusted classification score corresponding to a certain state category is divided by the total score value to obtain the confidence score value corresponding to that state category. This division process is repeated for all state categories to obtain multiple confidence score values, and the sum of all confidence score values ​​is made up to one.

[0082] Next, the maximum value is extracted by traversing multiple confidence scores to determine the target confidence score. Specifically, the first confidence score is read and recorded as the current maximum value. Then, the second to the last confidence scores are read sequentially. For each confidence score read, it is compared with the current maximum value. If the confidence score is greater than the current maximum value, the current maximum value is replaced with the confidence score. If the confidence score is less than or equal to the current maximum value, the current maximum value is kept unchanged. After all confidence scores have been traversed, the final retained current maximum value is determined as the target confidence score.

[0083] Next, the target confidence score value is matched with multiple state categories to determine the target state category, which is then used as candidate state data. In this process, the target confidence score value is first located within the multiple confidence scores, and then the state category matching that position is read from the multiple state categories and identified as the target state category. Subsequently, the target state category and the target confidence score value are stored accordingly, thus forming candidate state data that simultaneously contains the current judgment state and its corresponding confidence level.

[0084] Subsequently, hierarchical decision-making is performed using the candidate state data. Specifically, the target state category and target confidence score value are first read from the candidate state data, and then decision calculations are performed layer by layer according to the preset decision levels. In the first level, the corresponding disposal type is found based on the target state category; in the second level, the corresponding execution method is found based on the disposal type and target confidence score value; in the third level, the corresponding execution parameters are found based on the execution method; if there are subsequent levels, the decision-making is completed layer by layer by using the output results of the previous level as the input basis for the next level. The decision-making at each level can be performed using a lookup table method, that is, the correspondence between state category, confidence interval, disposal type, execution method, and execution parameters is pre-established, and the result of each level is obtained by matching the correspondence; after all levels are processed, multi-level auxiliary decision messages are obtained in hierarchical order.

[0085] Finally, the multi-layer auxiliary decision messages are associated and encapsulated. All levels of content in the multi-layer auxiliary decision messages are read, and then the target state category, target confidence score value, first-layer decision content, second-layer decision content, third-layer decision content, and subsequent-layer decision content are written sequentially according to the instruction output format of the intelligent fusion terminal. During the writing process, the decision content of the previous layer is arranged before the decision content of the next layer, and the order relationship between the layers remains unchanged. After all the content is written, a complete auxiliary decision instruction is formed, which enables the intelligent fusion terminal to directly execute the corresponding state response processing based on the auxiliary decision instruction.

[0086] In summary, the embodiments of this application have at least the following technical effects: This application utilizes an intelligent fusion terminal to perform real-time data acquisition and sensing of a target application scenario, obtaining a raw time-series data stream. Multimodal data fusion is then performed on the raw time-series data stream to construct a fused time-series data stream. A dual-layer feature extraction window is constructed, comprising a decay window and a dynamic sliding window mechanism. Real-time feature fusion extraction is performed on the fused time-series data stream using the decay window and the dynamic sliding window to obtain multi-scale real-time feature vectors. Based on the multi-scale real-time feature vectors, an initial feature vector matrix is ​​constructed and dynamically updated with feature weight allocation to obtain a feature vector matrix. The intelligent fusion terminal responds to the feature vector matrix with state recognition, generating auxiliary decision-making instructions for the intelligent fusion terminal. This invention addresses the technical problem of insufficient real-time feature extraction capability of existing intelligent fusion terminals for built-in time-series data. By constructing a dual-layer feature extraction window including a decay window and a dynamic sliding window, real-time feature fusion extraction is performed on the fused time-series data stream, thereby improving the real-time feature extraction effect of the intelligent fusion terminal for built-in time-series data.

[0087] Example 2, based on the same inventive concept as the real-time feature extraction method for built-in time-series data in the intelligent fusion terminal described in the previous examples, such as... Figure 2 As shown, this application provides a real-time feature extraction system for time-series data built into an intelligent fusion terminal. The system and method embodiments in this application are based on the same inventive concept. The system includes: The data acquisition module 11 is used to collect and sense the target application scenario in real time through the intelligent fusion terminal to obtain the raw time-series data stream; the data fusion module 12 is used to perform multimodal data fusion on the raw time-series data stream to construct a fused time-series data stream; the feature fusion extraction module 13 is used to construct a two-layer feature extraction window, the two-layer feature extraction window mechanism including a decay window and a dynamic sliding window, and to perform real-time feature fusion extraction on the fused time-series data stream through the decay window and the dynamic sliding window to obtain multi-scale real-time feature vectors; the weight allocation update module 14 is used to construct an initial feature vector matrix based on the multi-scale real-time feature vectors and perform dynamic feature weight allocation update to obtain a feature vector matrix; the instruction generation module 15 is used to perform state recognition response to the intelligent fusion terminal according to the feature vector matrix and generate auxiliary decision-making instructions for the intelligent fusion terminal.

[0088] Furthermore, the system is also used to implement the following functions: A reference time axis is constructed, and the original time-series data stream is subjected to deviation analysis according to the reference time axis to extract time offset data. The original time-series data stream is updated according to the time offset data and synchronized with the reference time axis to generate a time-series aligned dataset. The time-series aligned dataset is traversed to perform local data analysis to obtain local data distribution characteristics. Based on the local data distribution characteristics, instantaneous abnormal fluctuation analysis is performed to extract instantaneous abnormal jump points. The time-series aligned dataset is cleaned according to the instantaneous abnormal jump points to obtain a cleaned time-series dataset. The cleaned time-series dataset is traversed to mark gaps and determine multiple data gap positions. Based on the cleaned time-series dataset, an adjacency search is performed according to the multiple data gap positions to extract multiple valid sensor data points. The multiple valid sensor data points are filled into the multiple data gap positions to generate the fused time-series data stream.

[0089] Furthermore, the system is also used to implement the following functions: Based on the time-series aligned dataset, a local sliding analysis is performed according to the time series to define a local neighborhood analysis window. First starting position information and first ending position information are defined according to the time series based on the time series aligned dataset. The local neighborhood analysis window moves point-by-point from the first starting position information to the first ending position information, reading the set of sensor data point values. The set of sensor data point values ​​is sorted in descending order of value size to obtain a value sequence. The midpoint of the value sequence is located, and a first value is extracted as the local central trend representation value of the local neighborhood analysis window. Discrete analysis is performed on the set of sensor data point values ​​through the local neighborhood analysis window to calculate the numerical dispersion. The local central trend representation value is combined with the numerical dispersion for local analysis to obtain the local data distribution characteristics.

[0090] Furthermore, the system is also used to implement the following functions: Based on the cleaned time-series dataset, second start position information and second end position information are defined according to the time sequence. The local neighborhood analysis window sequentially traverses from the second start position information to the second end position information to determine if there are any empty space placeholders in the data storage locations of the cleaned time-series dataset. When an empty space placeholder exists in the data storage location of the cleaned time-series dataset, a record extraction instruction is generated to extract multiple data gaps to be filled. The multiple data gaps to be filled are then searched forward according to the time sequence to obtain a first valid sensor data point, which includes a first distance value between the first valid sensor data point and the target data gap to be filled. The system searches backward in time sequence according to the multiple data gaps to be filled, obtaining second valid sensor data points. Each second valid sensor data point contains a second distance value between itself and the target data gap. When the first distance value is greater than the second distance value, the second valid sensor data point is selected for data completion to generate a fused time-series data stream. When the first distance value is less than the second distance value, the first valid sensor data point is selected for data completion to generate a fused time-series data stream. When the first distance value is equal to the second distance value, either the first valid sensor data point or the second valid sensor data point is randomly selected for data completion to generate the fused time-series data stream.

[0091] Furthermore, the system is also used to implement the following functions: The memory buffer of the intelligent fusion terminal is accessed and defined to determine the historical data observation interval; historical time-series data segments are extracted based on the historical data observation interval; spatial capacity extreme value analysis is performed based on the memory buffer, and window capacity parameters are set according to the capacity extreme values; data attenuation weights are allocated to the attenuation window based on the historical time-series data segments to determine the attenuation weight coefficient; weighted feature calculation is performed on the fused time-series data stream through the attenuation window and the attenuation weight coefficient to generate a first-scale real-time feature vector.

[0092] Furthermore, the system is also used to implement the following functions: Based on the memory buffer, memory is allocated according to the dynamic sliding window to generate an independent memory address space; data containment analysis is performed based on the independent memory address space, and an initial window size parameter is set; based on the initial window size parameter, adjacent data in the fused time-series data stream are continuously read, and the numerical difference is calculated; the time sampling interval data of adjacent data is extracted, and the numerical difference is divided by the time sampling interval data to calculate the real-time change rate; based on the real-time change rate, data stream change analysis is performed to generate a data stream change trend, and the initial window size parameter of the dynamic sliding window is dynamically adjusted to generate a window size parameter; through the dynamic sliding window according to the window size parameter, real-time feature fusion extraction is performed on the fused time-series data stream to obtain a second-scale real-time feature vector.

[0093] Furthermore, the system is also used to implement the following functions: Dimensional analysis is performed based on the multi-scale real-time feature vectors to extract multiple feature dimensions. The number of historical features retained is retrieved from the intelligent fusion terminal, and this number is used as the number of rows, while the multiple feature dimensions are used as the number of columns to construct a two-dimensional array space. Data is filled into the two-dimensional array space to construct an initial feature vector matrix. A one-dimensional weight storage array is created based on the multiple feature dimensions, and weights are assigned to the initial feature vector matrix using this one-dimensional weight storage array to determine multiple initial weight arrays. The initial feature vector matrix is ​​traversed to calculate contribution, obtaining multiple contribution scores. The multiple initial weight arrays are iteratively updated according to the multiple contribution scores to generate multiple weight arrays. The initial feature vector matrix is ​​then reweighted according to the multiple weight arrays to generate the feature vector matrix.

[0094] Furthermore, the system is also used to implement the following functions: The data filling judgment is performed by traversing the two-dimensional array space: S1: Determine whether the number of rows filled with feature vectors in the two-dimensional array space has reached the number of historical features to be retained; S2: When the number of rows filled with feature vectors in the two-dimensional array space has not reached the number of historical features to be retained, the multi-scale real-time feature vectors are filled into the first empty row of the two-dimensional array space according to the multiple feature dimensions; S3: When the number of rows filled with feature vectors in the two-dimensional array space has reached the number of historical features to be retained, the row feature vector data with the longest storage time maximum value in the two-dimensional array space is removed as a whole, and the remaining rows are shifted forward one row to fill the empty space, obtaining empty row spaces; S4: The multi-scale real-time feature vectors are filled into the empty row spaces of the two-dimensional array space; Based on the above steps S1, S2, S3, and S4, continuous filling is performed to construct the initial feature vector matrix.

[0095] Furthermore, the system is also used to implement the following functions: The feature vector matrix is ​​synchronized to the state classifier built into the intelligent fusion terminal for analysis to generate multiple state categories; confidence assessment is performed based on the multiple state categories to generate multiple confidence scores; the maximum value is extracted from the multiple confidence scores to determine the target confidence score; the target confidence score is matched with the multiple state categories to determine the target state category, and the target state category table is used as candidate state data; hierarchical decision-making is performed using the candidate state data to obtain multi-level auxiliary decision messages; the multi-level auxiliary decision messages are associated and encapsulated to construct auxiliary decision instructions.

[0096] It should be noted that the order of the embodiments described above is for descriptive purposes only and does not represent the superiority or inferiority of the embodiments. Furthermore, the above description focuses on specific embodiments of this specification. The processes depicted in the accompanying drawings do not necessarily require a specific or sequential order to achieve the desired results. In some implementations, multitasking and parallel processing are possible or may be advantageous.

[0097] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any modifications, equivalent changes, and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A method for real-time feature extraction of time-series data built into an intelligent fusion terminal, characterized in that, The method includes: The target application scenario is collected and sensed in real time by intelligent fusion terminal to obtain raw time-series data stream; Multimodal data fusion is performed on the original time-series data stream to construct a fused time-series data stream; A dual-layer feature extraction window is constructed, the dual-layer feature extraction window mechanism includes a decay window and a dynamic sliding window, and the fused time-series data stream is subjected to real-time feature fusion extraction through the decay window and the dynamic sliding window to obtain multi-scale real-time feature vectors; Based on the multi-scale real-time feature vectors, an initial feature vector matrix is ​​constructed, and dynamic feature weight allocation and updating are performed to obtain the feature vector matrix. Based on the feature vector matrix, the intelligent fusion terminal performs state recognition response and generates auxiliary decision-making instructions for the intelligent fusion terminal.

2. The real-time feature extraction method for built-in time-series data in an intelligent fusion terminal as described in claim 1, characterized in that, The method for performing multimodal data fusion on the original time-series data stream to construct a fused time-series data stream includes: A reference time axis is constructed, and the original time series data stream is subjected to deviation analysis according to the reference time axis to extract time offset data. The original time-series data stream is updated according to the time offset data and synchronized with the reference time axis to generate a time-series aligned dataset. The time-aligned dataset is traversed to perform local data analysis and obtain local data distribution characteristics. Based on the local data distribution characteristics, instantaneous abnormal fluctuation analysis is performed to extract instantaneous abnormal jump points; Based on the instantaneous abnormal jump points, the time-series aligned dataset is cleaned to obtain a cleaned time-series dataset; The cleaned time-series dataset is traversed to identify gaps and determine multiple data gap locations. Based on the cleaned time-series dataset, an adjacency search is performed according to the multiple data gap locations to extract multiple valid sensor data points. The multiple valid sensor data points are then filled into the multiple data gap locations to generate the fused time-series data stream.

3. The real-time feature extraction method for built-in time-series data in an intelligent fusion terminal as described in claim 2, characterized in that, The method involves traversing the time-aligned dataset to perform local data analysis and obtain local data distribution characteristics. Based on the time-series aligned dataset, a local sliding analysis is performed according to the time series, and a local neighborhood analysis window is defined. Based on the time-series aligned dataset, the first start position information and the first end position information are defined according to the time sequence. The local neighborhood analysis window moves point by point from the first starting position information to the first ending position information to read the set of sensor data point values. The set of sensor data points is sorted in descending order of numerical value to obtain a numerical sequence; The midpoint of the numerical sequence is located, and the first value is extracted as the local central trend characterization value of the local neighborhood analysis window; The local neighborhood analysis window is used to perform discrete analysis on the set of sensor data points and calculate the numerical dispersion. By combining the local central tendency characterization value with the numerical dispersion, local analysis is performed to obtain the local data distribution characteristics.

4. The real-time feature extraction method for built-in time-series data in an intelligent fusion terminal as described in claim 3, characterized in that, The method involves traversing the cleaned time-series dataset to identify gaps, determining multiple data gap locations, performing adjacency searches based on these gap locations to extract multiple valid sensor data points, and filling these valid sensor data points into the gap locations to generate the fused time-series data stream. Based on the cleaning time-series dataset, define the second start position information and the second end position information according to the time sequence. The local neighborhood analysis window is used to sequentially traverse from the second starting position information to the second ending position information to determine whether there is a space placeholder identifier in the data storage location of the cleaned time series dataset. When the empty space placeholder identifier exists in the data storage location of the cleaning time series dataset, a record extraction instruction is generated to extract multiple data gaps to be filled. Searching forward according to the time sequence based on the multiple data gap locations to be filled, a first effective sensing data point is obtained. The first effective sensing data point includes a first distance value between itself and the target data gap location to be filled. Search backwards in time sequence according to the multiple data gap locations to be filled to obtain a second effective sensing data point, wherein the second effective sensing data point contains a second distance value between the target data gap location to be filled; When the first distance value is greater than the second distance value, the second valid sensor data point is selected to complete the data and generate a fused time-series data stream; When the first distance value is less than the second distance value, the first valid sensor data point is selected to complete the data and generate a fused time-series data stream. When the first distance value is equal to the second distance value, the first effective sensor data point or the second effective sensor data point is randomly selected to complete the data and generate the fused time-series data stream.

5. The real-time feature extraction method for built-in time-series data in an intelligent fusion terminal as described in claim 1, characterized in that, Real-time feature fusion and extraction are performed on the fused time-series data stream through the attenuation window to obtain multi-scale real-time feature vectors. The method includes: The memory buffer of the intelligent fusion terminal is accessed and defined to determine the historical data observation interval; Historical time-series data segments are extracted based on the historical data observation intervals; Based on the memory buffer, perform space capacity extreme value analysis, and set window capacity parameters according to the capacity extreme values; Based on the historical time-series data segments, the attenuation window is weighted to determine the attenuation weight coefficient. The fused time-series data stream is weighted and its features are calculated using the attenuation window and the attenuation weight coefficient to generate a first-scale real-time feature vector.

6. The real-time feature extraction method for built-in time-series data in an intelligent fusion terminal as described in claim 5, characterized in that, The method involves performing real-time feature fusion and extraction on the fused time-series data stream using the dynamic sliding window to obtain multi-scale real-time feature vectors. Based on the memory buffer, memory is allocated according to the dynamic sliding window to generate an independent memory address space; Data containment analysis is performed based on the independent memory address space, and initial window size parameters are set. Based on the initial window size parameter, the fused time-series data stream is continuously read from adjacent data to calculate the numerical difference. Extract the time sampling interval data of adjacent data, divide the numerical difference by the time sampling interval data, and calculate the real-time rate of change; Based on the real-time change rate, data stream change analysis is performed to generate a data stream change trend. The initial window size parameters of the dynamic sliding window are then dynamically adjusted to generate window size parameters. The dynamic sliding window is used to perform real-time feature fusion and extraction on the fused time-series data stream according to the window size parameters to obtain a second-scale real-time feature vector.

7. The real-time feature extraction method for built-in time-series data in an intelligent fusion terminal as described in claim 1, characterized in that, Based on the multi-scale real-time feature vectors, an initial feature vector matrix is ​​constructed, and dynamic feature weight allocation and updating are performed to obtain the feature vector matrix. The method includes: Dimensional analysis is performed based on the multi-scale real-time feature vectors to extract multiple feature dimensions; Based on the intelligent fusion terminal, the number of historical features retained is retrieved, and the number of historical features retained is used as the number of rows, and the multiple feature dimensions are used as the number of columns to construct a two-dimensional array space; The two-dimensional array space is filled with data to construct an initial feature vector matrix; A one-dimensional weight storage array is created based on the multiple feature dimensions, and the initial feature vector matrix is ​​weighted using the one-dimensional weight storage array to determine multiple initial weight arrays. The initial feature vector matrix is ​​traversed to calculate the contribution, resulting in multiple contribution scores. The multiple initial weight arrays are iteratively updated based on the multiple contribution scores to generate multiple weight arrays; The initial feature vector matrix is ​​reweighted according to the multiple weight arrays to generate the feature vector matrix.

8. The real-time feature extraction method for built-in time-series data in an intelligent fusion terminal as described in claim 7, characterized in that, The method for filling the two-dimensional array space with data to construct an initial feature vector matrix includes: Iterate through the two-dimensional array space to determine data filling: S1: Determine whether the number of rows in the two-dimensional array space that have been filled with feature vectors has reached the number of historical features to be retained; S2: When the number of rows in the two-dimensional array space that have been filled with feature vectors has not reached the number of historical features to be retained, the multi-scale real-time feature vectors are filled into the first empty row of the two-dimensional array space according to the multiple feature dimensions. S3: When the number of rows in the two-dimensional array space that have been filled with feature vectors reaches the number of historical features to be retained, the row feature vector data with the longest storage time in the two-dimensional array space is removed as a whole, and the remaining rows are shifted forward one row to fill the empty space, thus obtaining empty rows in the space. S4: Fill the empty spaces in the two-dimensional array space with the multi-scale real-time feature vector; Based on the above steps S1, S2, S3, and S4, the initial feature vector matrix is ​​continuously filled to construct it.

9. The real-time feature extraction method for built-in time-series data in an intelligent fusion terminal as described in claim 1, characterized in that, The method includes performing state recognition response on the intelligent fusion terminal based on the feature vector matrix, and generating auxiliary decision-making instructions for the intelligent fusion terminal. The feature vector matrix is ​​synchronized to the state classifier built into the intelligent fusion terminal for analysis to generate multiple state categories; Confidence assessment is performed based on the multiple state categories to generate multiple confidence score values; The maximum value is extracted by traversing the multiple confidence score values ​​to determine the target confidence score value; The target state category is determined by matching the target confidence score value with the multiple state categories, and the target state category table is used as candidate state data. Hierarchical decision-making is performed using the candidate state data to obtain multi-level auxiliary decision messages; The multi-layered auxiliary decision messages are associated and encapsulated to construct auxiliary decision instructions.

10. A real-time feature extraction system for time-series data is built into the intelligent fusion terminal, characterized in that: The system is used to execute the real-time feature extraction method for time-series data built into the intelligent fusion terminal as described in any one of claims 1-9, and the system includes: The data acquisition module is used to collect and sense data from the target application scenario in real time through an intelligent fusion terminal to obtain raw time-series data streams. The data fusion module is used to perform multimodal data fusion on the original time-series data stream to construct a fused time-series data stream; The feature fusion and extraction module is used to construct a two-layer feature extraction window. The two-layer feature extraction window mechanism includes a decay window and a dynamic sliding window. The decay window and the dynamic sliding window are used to perform real-time feature fusion and extraction on the fused time-series data stream to obtain multi-scale real-time feature vectors. The weight allocation and update module is used to construct an initial feature vector matrix based on the multi-scale real-time feature vectors and perform dynamic feature weight allocation and update to obtain the feature vector matrix. The instruction generation module is used to perform state recognition response on the intelligent fusion terminal based on the feature vector matrix and generate auxiliary decision-making instructions for the intelligent fusion terminal.

Images