CN115840877B - Distributed stream processing method, system, storage medium and computer extracted from MFCC - Google Patents

Abstract

The present invention relates to the distributed stream processing method, system, storage medium and computer that MFCC extracts, and its method comprises the following steps: obtain multi-source signal original data stream in parallel; The data type of multi-source signal original data stream is String data; Multi-source Perform parallel flat mapping on the original signal data stream to obtain multi-source discrete signal data streams; perform data stream windowing operation on multi-source discrete signal data streams to obtain parallel continuous sliding windows; use parallel window processing functions in parallel continuous Mel frequency cepstral coefficients are extracted in the sliding window of the multi-source signal to obtain the corresponding Mel frequency cepstral coefficient data stream of the multi-source signal; the present invention makes the Mel frequency of the multi-source data stream The extraction of cepstral coefficients can be performed in parallel, improving the efficiency and timeliness of the extraction of Mel-frequency cepstral coefficients, and avoiding the lag of offline processing of a large number of data and the pressure of data processing caused by the huge amount of data processing.

Description

Distributed stream processing method, system, storage medium and computer extracted from MFCC

technical field

The invention relates to the field of industrial big data and signal processing, and is used for real-time processing of mechanical equipment vibration signals and sound signals, in particular to a distributed stream processing method, system, storage medium and computer extracted by MFCC.

Background technique

MFCC, the full name is Mel Frequency Cepstrum Coefficient, which is the Mel frequency cepstrum coefficient. MFCC is often used in the processing of sound signals and also in vibration signal processing. In the operation and maintenance scenario of mechanical equipment, vibration monitoring and sound monitoring are commonly used technical means. Through the MFCC that extracts the signal, it is used for abnormal detection and fault diagnosis of mechanical equipment.

The currently known MFCC extraction methods and technologies all perform off-line analysis after the vibration data or sound data collection is completed, and the MFCC extraction algorithm needs to process the collected signals in frames. When the number of monitored devices increases and the number of vibration and sound sensors becomes larger and larger, the signal will form a huge matrix after being framed, which will bring huge pressure to the offline feature extraction work, and due to the lag of feature extraction, it is difficult to meet Online fault diagnosis scenarios of MFCCs need to be analyzed in real time.

Contents of the invention

In order to solve the problem that when the number of monitored devices is large and the number of vibration and sound sensors becomes larger and larger, the signal will form a huge matrix after being framed, which will bring huge pressure to the offline feature extraction work, and because the feature extraction lags behind, It is difficult to meet technical problems such as real-time analysis of MFCC online fault diagnosis scenarios, and the present invention provides a distributed stream processing method, system, storage medium and computer for MFCC extraction.

The technical scheme that the present invention solves the problems of the technologies described above is as follows:

The distributed stream processing method extracted by MFCC includes the following steps:

Obtaining multi-source signal original data streams in parallel; wherein, the data type of the multi-source signal original data streams is String data;

performing parallel flat mapping on the multi-source signal original data stream to obtain a multi-source discrete signal data stream;

Performing a data stream windowing operation on the multi-source discrete signal data stream to obtain parallel continuous sliding windows;

The parallel window processing function is used to extract the Mel-frequency cepstral coefficients in the parallel continuous sliding windows, so as to obtain the Mel-frequency cepstral coefficient data streams corresponding to the multi-source signals.

The beneficial effects of the present invention are: vibration, sound and other signals generate a lot of data points per millisecond, if the data points are sent one by one, it is very likely that the data points that occur first will arrive at the processing system. The data received by the system will be out of order. Therefore, the present invention encapsulates the data collected in multiple milliseconds into a segment in String format to send, and the signal data points in the segment are arranged according to the original order of occurrence, so they will not be out of sequence. Milliseconds, there will be no out-of-order between fragments when the network transmission is normal. When there are multiple data sources of the original data stream, the extraction of Mel-frequency cepstral coefficients of multi-source data streams can be performed in parallel by performing flat mapping and partitioning and windowing operations on the data streams, improving the Mel-frequency cepstral coefficients The extraction efficiency and timeliness are high, avoiding the lag of offline processing of a large amount of data and the data processing pressure caused by the huge amount of data processing.

On the basis of the above technical solutions, the present invention can also be improved as follows.

Further, the String data includes at least a timestamp, a sensor ID, a signal value, and a delimiter, wherein the sensor ID is a sensor number corresponding to the original data stream.

Further, performing parallel flat mapping on the multi-source original data stream to obtain a multi-source discrete signal data stream includes the following steps: performing parallel flat mapping on the multi-source original data stream by using the Flink stream processing method to obtain a multi-source discrete signal data flow.

The beneficial effect of adopting the above further solution is that Flink stream processing has the characteristics of high throughput, low latency, and distributed, which improves the data processing efficiency.

Further, performing a data stream windowing operation on the multi-source discrete signal data stream to obtain parallel continuous sliding windows includes the following steps:

Performing a keyBy operation on the multi-source discrete signal data stream according to the sensor ID to obtain a keyed data stream; wherein, the keyBy operation is specifically sending the multi-source discrete signal data stream with the same sensor ID to a designated partition;

A data stream windowing operation is performed on the keyed data stream in each partition to obtain parallel and continuous sliding windows corresponding to each sensor.

Further, using the parallel window processing function to extract the Mel-frequency cepstral coefficients in the continuous sliding window, to obtain the Mel-frequency cepstral coefficient data stream, comprises the following steps:

storing the data in each of the sliding windows in a corresponding double-precision array by using the parallel window processing function;

Calling a Mel-frequency cepstral coefficient extraction function for each of the double-precision arrays to obtain the Mel-frequency cepstral coefficient data stream.

Further, the Mel-frequency cepstral coefficient extraction function includes a main function and a plurality of sub-functions, and the plurality of sub-functions are respectively a Mel filter bank function, a discrete cosine transform function, a fast Fourier transform function, and a Hamming window function;

The double-precision array is input into the main function, and the main function calculates the double-precision array by calling a plurality of the sub-functions to obtain Mel-frequency cepstral coefficients.

In order to solve the above-mentioned technical problems, the present invention also provides a distributed stream processing system for extracting multi-source signal Mel-frequency cepstral coefficients, the specific technical content of which is as follows:

The distributed stream processing system extracted by MFCC, including:

A data acquisition module, configured to acquire multi-source signal original data streams in parallel; wherein, the data type of the multi-source signal original data streams is String data;

The data processing module is used to perform parallel flat mapping on the multi-source signal original data stream to obtain a multi-source discrete signal data stream; perform a data stream windowing operation on the multi-source discrete signal data stream to obtain parallel continuous Sliding window: using a parallel window processing function to extract Mel-frequency cepstral coefficients in the continuous sliding windows to obtain a data stream of Mel-frequency cepstral coefficients corresponding to multi-source signals.

Based on the distributed stream processing method for extracting multi-source signal Mel-frequency cepstral coefficients, the present invention also provides a storage medium, the technical content of which is as follows:

A storage medium stores a computer program, and when the computer program is executed by a computer processor, the above-mentioned distributed flow processing method for extracting Mel-frequency cepstral coefficients of multi-source signals is realized.

Based on the multi-source signal Mel frequency cepstral coefficient distributed stream processing extraction method, the present invention also provides a computer, the technical content of which is as follows:

A computer, including a memory and a processor, the memory stores a computer program, and when the computer program is executed by the processor, the above-mentioned distributed flow processing method extracted by MFCC is realized.

Description of drawings

Fig. 1 is the block flow diagram of the distributed stream processing method that a kind of MFCC extracts in the embodiment of the present invention 1;

FIG. 2 is a schematic diagram of a signal flow flat mapping processing flow in Embodiment 1 of the present invention;

Fig. 3 is the structural representation of MFCC extraction function set in the embodiment 1 of the present invention;

Fig. 4 is the graph of original vibration signal in the embodiment 3 of the present invention;

Fig. 5 is the vibration signal data flow recording figure in the embodiment 3 of the present invention;

Fig. 6 is a block diagram of the execution of the Flink MFCC extraction task program in Embodiment 3 of the present invention.

Detailed ways

The principles and features of the present invention are described below in conjunction with the accompanying drawings, and the examples given are only used to explain the present invention, and are not intended to limit the scope of the present invention.

Example 1

As shown in Figure 1, the present embodiment provides a distributed stream processing method extracted by MFCC, including the following steps:

S1. Obtain multi-source signal original data streams in parallel; wherein, the data type of the original data stream is String data; wherein, the String data includes at least a timestamp, a sensor ID, a signal value, and a separator, and the sensor ID is The sensor number corresponding to the raw data stream. For vibration signal and sound signal data sources, the frequency of signal acquisition is often high. Taking the collection frequency of 20kHz as an example, 20 points will be generated every millisecond. In order to avoid the out-of-order problem caused by the signal transmission through the network, Kafka producers send message records in a way of cyclically packing several milliseconds. Kafka is an open source stream processing platform developed by the Apache Software Foundation. It is a high-throughput distributed publish-subscribe messaging system written in Scala and Java.

S2. Perform parallel flat mapping of the multi-source signal original data stream to obtain a multi-source discrete signal data stream; specifically, use the Flink stream processing method to perform parallel flat mapping on the multi-source signal original data stream to obtain a multi-source discrete signal data stream. Signal data flow; in big data scenarios, Kafka producers are used as data sources, and Flink programs pull and consume data from Kafka.

S3. The multi-source discrete signal data stream performs a data stream windowing operation on the multi-source discrete signal data stream to obtain parallel continuous sliding windows; the specific steps are: divide the multi-source discrete signal data stream according to the sensor The ID performs a keyBy operation to obtain a keyed data stream; wherein, the keyBy operation is specifically to send the multi-source discrete signal data stream with the same sensor ID to the specified same partition;

A data stream windowing operation is performed on the keyed data stream in each specified partition to obtain parallel continuous sliding windows.

S4. Using a parallel window processing function to extract Mel-frequency cepstral coefficients in the parallel continuous sliding windows to obtain a data stream of Mel-frequency cepstral coefficients corresponding to multi-source signals. The specific steps are:

storing the data in each of the sliding windows in a corresponding double-precision array by utilizing a parallel window processing function;

calling the Mel-frequency cepstral coefficient extraction function for each of the double-precision arrays to obtain the Mel-frequency cepstral coefficient data stream;

Wherein, the Mel-frequency cepstral coefficient extraction function includes a main function and a plurality of sub-functions, and the plurality of sub-functions are respectively a Mel filter bank function, a discrete cosine transform function, a fast Fourier transform function, and a Hamming window function;

The double-precision array is input into the main function, and the main function calculates the double-precision array by calling a plurality of the sub-functions to obtain Mel-frequency cepstral coefficients. Call the Mel filter bank generation function to construct the Mel filter bank; then, call the DCT coefficient matrix generation function to obtain the DCT coefficients; then carry out fast Fourier transform to the double-precision array to obtain the spectrum result; finally, the spectrum The result is filtered, that is, after the Mel filter bank is multiplied by the spectral result, the logarithm is calculated and multiplied by the DCT coefficient to obtain the Mel frequency cepstral coefficient.

As shown in Figure 2, the original data of the multi-source signal is sent by the corresponding Kafka partition, so that the Flink stream processing program can be used to read data from the data source in parallel, and the original data stream formed after the data is read is a String Format data, and then need to flat map the data in String format. Flink's flapMap operator is used to implement this conversion operation, and the core is the custom FlatMapFunction. The main steps of the flapMap operator are: divide the records in the String format into a string array Array[String] according to the delimiter; The element starts with the first signal value, and each cycle sends out data packaged in the following caseclass format:

case class VibElement(ts:Long,sensorId:Int,acc:Double)

Among them, ts is the timestamp of the original record, sensorId is the sensor ID of the original record, and acc is the discrete signal value.

As shown in Figure 3, the keyBy operation is performed on the data stream obtained after the flat mapping conversion according to the sensor ID to form a keyed data stream KeyedStream, and subsequent windowing and window processing are performed independently and in parallel for different sensor data streams. Among them, the main function of keyBy is to send the data with the same key to the same partition; the data is originally distributed in different slots or partitions, and keyBy will pull the data of the same key to the same slot or partition.

The windowing operation adopts a counting sliding window, the window length is the number of data elements, and the sliding step is the two parameters of the windowing operation. In signal MFCC extraction applications, the window length is generally set to 256, and the sliding step is set to 128.

When all the data is collected for each window, store the data in an array of Double type, and then call the MFCC extraction function on the array of Double type to obtain the MFCC of DenseVector[Double] type. In order to identify the sensor ID to which the MFCC data belongs and the timestamp of the signal window, the final extraction result is packaged into the following type of sample class output:

case class MfccResult(startts: Long, endts: Long,

sensorId:Int,mfcc:DenseVector[Double])

Among them, startts is extracted from the first piece of data in the window, which is the window start time stamp, endts is extracted from the last piece of data in the window, and is the window end time stamp, sensorId is the key tag carried by the keyed data stream . Here, the time stamps are event time, that is, the acquisition time of the sensor signal.

As shown in Figure 4, Scala is a commonly used development language for big data technology. In order to be directly called by the Flink main program for signal flow feature extraction, the MFCC extraction function set developed with Scala is provided. The MFCC extraction function set includes the main function and Mel filter. Group generation function, DCT function, FFT function, Hamming window function. The DCT function is the above discrete cosine transform coefficient matrix generation function, the FFT function is the fast Fourier transform function, the Hamming window function is the Hamming window function, the Mel filter bank generation function is the Mel filter bank generation function, and the Mel filter The group is a Mel filter bank.

The flow, input and output, and function call relationship of the main function are as follows:

①The design of the main function:

The input of the main function is the array x of the Array[Double] type, and also includes the sampling rate fs, the Mel filter order p, and the FFT transformation length N. The data types of these three parameters are Int, that is, the integer data type, Where p is generally set to 24. FFT means Fast Fourier Transform and the array x means the above array of doubles.

The output of the main function is MFCC, and the format of MFCC is the DenseVector[Double] type of p/2.

The main flow of the main function is: first, call the Mel filter bank generation function to construct the Mel filter bank; then, call the DCT coefficient matrix generation function to obtain the DCT coefficients; then perform fast Fourier transform on the array x to obtain the spectrum Result; finally, filter the spectrum result, that is, multiply the spectrum by the filter bank, calculate the logarithm and multiply it by the DCT coefficient to obtain the MFCC coefficient.

②Mel filter bank generation function, the three parameters fs, p, N input from the main function are passed to this function, adopt the conventional Mel filter bank generation method to obtain the Mel filter bank of DenseMatrix[Double] type, The dimensions of this matrix are:

p×(N/2+1)

Among them, the DCT coefficient matrix generation function is used to use the parameter p input from the main function as the input of the function, and the DCT coefficient matrix of the DenseMatrix[Double] type is obtained by using the conventional DCT coefficient generation method, and the size of the matrix is p/ 2 x p. Wherein, the DCT coefficients are discrete cosine transform coefficients.

The FFT function is used to take the array x input from the main function as the input of the function, and use the conventional FFT transformation method to obtain the output of the Array[Double] type. In the FFT function, the Hammi ng window filter is first performed on x.

The Hamming window filter function is used to take the array x input from the FFT function as the input of the function, and use the conventional Hamming window construction method to obtain an array of the same length as the filtered result.

In the embodiment of the present invention, by setting the data of the original data stream as String data, vibration, sound and other signals, many data points are generated every millisecond. If the data points are sent one by one, it is very likely that the first occurrence will occur A situation where the data points arrive at the processing system later, so that the data received by the processing system is out of order. Therefore, the present invention encapsulates the data collected in multiple milliseconds into a segment in String format to send, and the signal data points in the segment are arranged according to the original order of occurrence, so they will not be out of sequence. Milliseconds, there will be no out-of-order between fragments when the network transmission is normal. When there are multiple data sources of the original data stream, by performing flat mapping and windowing operations on the data stream, the extraction of the Mel-frequency cepstral coefficients of the multi-source data stream can be performed in parallel, and the efficiency of the Mel-frequency cepstral coefficients can be improved. The extraction efficiency and timeliness avoid the lag of offline processing of large amounts of data and the pressure of data processing caused by the huge amount of data processing. Flink stream processing has the characteristics of high throughput, low latency, and distribution, which improves the efficiency of data processing. Different from the offline feature extraction of Mel-frequency cepstral coefficients in the prior art, the embodiment of the present invention extracts Mel-frequency cepstral coefficients under the stream processing paradigm, and does not require frame division operation during the extraction process.

Example 2

Based on embodiment 1, this embodiment provides a distributed stream processing system extracted by MFCC, including a data acquisition module and a data processing module;

The data acquisition module is used to acquire multi-source signal original data flow in parallel; wherein, the data type of said original data flow is String data; the specific data acquisition method is: adopt Kafka producer as data source, continuously send the multi-source signal of String format The original data stream, and the Flink program pulls and consumes data from Kafka.

The data processing module is used to perform parallel flat mapping on the multi-source signal original data stream to obtain a multi-source discrete signal data stream; perform a data stream windowing operation on the multi-source discrete signal data stream to obtain parallel continuous Sliding window: using a parallel window processing function to extract Mel-frequency cepstral coefficients in the parallel continuous sliding windows to obtain a data stream of Mel-frequency cepstral coefficients corresponding to multi-source signals.

Wherein, the original data stream is subjected to parallel flat mapping to obtain a multi-source discrete signal data stream; specifically, the original data stream is subjected to parallel flat mapping by using the flatMap operator of Flink stream processing to obtain a multi-source discrete signal data stream .

The multi-source discrete signal data stream performs a data stream windowing operation on the multi-source discrete signal data stream to obtain parallel continuous sliding windows; the specific steps are: perform the multi-source discrete signal data stream according to the sensor ID keyBy operation to obtain the keyed data stream; wherein, the keyBy operation is specifically to send the multi-source discrete signal data stream with the same sensor ID to the same designated partition;

A data stream windowing operation is performed on the keyed data stream in each specified partition to obtain parallel continuous sliding windows.

The parallel window processing function is used to extract the Mel-frequency cepstral coefficients in the parallel continuous sliding windows, so as to obtain the Mel-frequency cepstral coefficient data streams corresponding to the multi-source signals. The specific steps are:

storing data in each of the sliding windows in a corresponding double-precision array using a window function;

calling the Mel-frequency cepstral coefficient extraction function for each of the double-precision arrays to obtain the Mel-frequency cepstral coefficient data stream;

Wherein, the Mel-frequency cepstral coefficient extraction function includes a main function and a plurality of sub-functions, and the plurality of sub-functions are respectively a Mel filter bank function, a discrete cosine transform function, a fast Fourier transform function, and a Hamming window function;

The double-precision array is input into the main function, and the main function calculates the double-precision array by calling a plurality of the sub-functions to obtain Mel-frequency cepstral coefficients.

Example 3

Based on Embodiment 1, this embodiment provides an experimental verification process and verification results of a distributed stream processing method extracted from MFCC or a distributed stream processing system extracted from MFCC. The specific experimental process and verification results are as follows:

The method and system proposed by the present invention are verified by taking the acceleration signal collected by the vibration sensor as an example. In this example, a total of 4 vibration sensors are arranged to collect the vibration signal of the equipment in real time. The sampling frequency of the signal is 20kHz. Figure 4 is a graph of the vibration signal within 1s.

The signal streams of 4 sensors are published in parallel through the Kafka producer, and a record is generated every 8ms. The record contains 160 sampling points, and the data stream record shown in Figure 5 is obtained.

Test items and test methods;

Four vibration signal data sources were used to test the MFCC extraction stream processing method and system. The main test items and their test methods are shown in Table 1:

Table 1 Test items and test methods

Test process and results;

MFCC extracts the distributed stream processing function test as follows:

Run the Kafka producer program and send the vibration signal data collected by 4 vibration sensors in parallel. Each sensor sends 160 sampling points every 8ms. The sampling points sent by each sensor every 8ms are represented as 1 record, and the continuous sending is 20,000 times. Record, the duration of the data flow is about 2.7 minutes, test whether it can be extracted normally and output the data flow of MFCC extraction results during this period.

According to the counting window with a length of 256 and a sliding step of 128, each sensor has 3,200,000 vibration sampling points, and the accumulated number of extracted MFCC results is:

The test results show:

Data stream generation and MFCC feature extraction are kept in sync, and the moment the signal is sent to the processing system, the MFCC feature extraction calculation task is triggered and completed; after repeated tests, the feature extraction results are completely consistent with the offline processing results, indicating that the program design and operation The correctness; the number of MFCC extraction results for each sensor is 24999, which proves that the integrity of data processing is 100%.

MFCC feature extraction delay time test:

In order to test the feature extraction delay time, run the Kafka producer program and the Flink feature extraction stream processing main program on the same host, and subtract the event time of the window cut-off from the instantaneous computer system time through MFCC feature processing to obtain the MFCC generated each time Delay time, a total of 99996 delay time data samples of four sensor feature extractions are obtained, and the relevant statistical results are shown in Table 2.

Table 2 Delay time test results

Since the test data is sent from the Kafka producer program running on the local host to the Kafka cluster, and then the Flink main program running on the local host pulls data from the Kafka cluster. After actual measurement, the processing time of the feature extraction itself for each window is very short and less than 1ms; therefore, most of the delay time is introduced by network transmission, even so the average overall delay of more than 30 milliseconds fully proves that the extraction through stream processing The MFCC feature is very efficient, that is to say, the delay time between the generation of the original data of the multi-source signal and the output of the corresponding MFCC feature is extremely short.

Test of the feature extractor on a computer cluster:

As shown in Figure 6, the FlinkMFCC extraction program and its third-party dependencies are packaged and deployed to the Hadoop cluster. Open a Yarn session for executing Flink tasks through the "./bin/yarn-session.sh-nmmfccflinktest-d" command. Then submit the Flink task through the "./bin/flinkrun-corg.atcsu.mfcc.VibMFCCRealTime/opt/program/FlinkMFCC-1.0.0.jar" command. Run the Kafka producer program to generate raw data of multi-source vibration signals. In cluster mode, the program runs normally, and the real-time extraction results of MFCC are correct.

The embodiment of the present invention has verified that the test results are consistent with the local stand-alone test results through experiments, showing that the original data stream of multi-source signals can be correctly extracted from the MFCC feature under the computer cluster processing mode, and the data processing is completed when the data is generated, and the data processing integrity rate is also 100%. Therefore, the present invention enables the extraction of Mel-frequency cepstral coefficients of multi-source data streams to be performed in real time and in parallel, improves the extraction efficiency and timeliness of Mel-frequency cepstral coefficients, and avoids the hysteresis and data The data processing pressure brought by the huge amount of processing.

Example 4

Based on Embodiment 1, this embodiment provides a storage medium, the storage medium stores a computer program, and when the computer program is executed by a processor of a computer, the above-mentioned distributed stream processing method extracted by MFCC is implemented. A storage medium refers to a carrier that stores data. Such as floppy disk, CD, DVD, hard disk, flash memory, U disk, CF card, SD card, MMC card, SM card, Memory Stick (Memory Stick), xD card, etc. The storage medium can also be based on flash memory, namely Nandflash, such as U disk, CF card, SD card, SDHC card, MMC card, SM card, memory stick, xD card, etc.

The present invention implements a distributed stream processing method that can implement the above-mentioned MFCC extraction by storing the program in a storage medium and executing the program through a processor, thereby improving data processing efficiency.

Example 5

Based on Embodiment 1, this embodiment provides a computer, including a memory and a processor, the memory stores a computer program, and when the computer program is executed by the processor, the above-mentioned distributed flow extracted by the MFCC is realized. Approach. By running or executing software programs and/or modules stored in the memory, and calling data stored in the memory, executing various functions of the terminal and processing data, the terminal is monitored as a whole, such as realizing the above-mentioned distributed MFCC extraction stream processing method. There may be one or more processors, and the processors may also be implemented as a combination of computing devices.

The embodiment of the present invention improves data processing efficiency by using a computer to implement the program or application module corresponding to the distributed stream processing method extracted by the MFCC.

The above descriptions are only preferred embodiments of the present invention, and are not intended to limit the present invention. Any modifications, equivalent replacements, improvements, etc. made within the concept and principles of the present invention shall be included in the protection of the present invention. within range.

Claims (7)

1. The distributed stream processing method that MFCC extracts is characterized in that, comprises the steps: Obtaining multi-source signal original data streams in parallel; wherein, the data type of the multi-source signal original data streams is String data; performing parallel flat mapping on the multi-source signal original data stream to obtain a multi-source discrete signal data stream; Performing a data stream windowing operation on the multi-source discrete signal data stream to obtain parallel continuous sliding windows; Using a parallel window processing function to extract Mel-frequency cepstral coefficients in parallel continuous sliding windows to obtain a Mel-frequency cepstral coefficient data stream corresponding to multi-source signals; Utilize the parallel window processing function to extract the Mel-frequency cepstral coefficients in the continuous sliding window, and obtain the Mel-frequency cepstral coefficient data flow, comprising the following steps: storing the data in each of the sliding windows in a corresponding double-precision array by using the parallel window processing function; calling the Mel-frequency cepstral coefficient extraction function for each of the double-precision arrays to obtain the Mel-frequency cepstral coefficient data stream; The Mel frequency cepstral coefficient extraction function includes a main function and a plurality of sub-functions, and a plurality of sub-functions are respectively a Mel filter bank function, a DCT function, a Fast Fourier Transform FFT function and a Hamming window function; The double-precision array is input into the main function, and the main function calculates the double-precision array by calling a plurality of the sub-functions to obtain Mel-frequency cepstral coefficients. 2. the distributed stream processing method that MFCC according to claim 1 extracts is characterized in that, described String data comprises timestamp, sensor ID, a plurality of signal values and delimiter at least, and wherein, described sensor ID is all The sensor number corresponding to the original data stream of the multi-source signal. 3. the distributed stream processing method that MFCC according to claim 2 extracts is characterized in that, described multi-source signal original data stream is carried out parallel flat mapping, obtains multi-source discrete signal data stream, comprises the steps: utilize Flink The stream processing method performs a parallel flat mapping operation on the original data stream to obtain a multi-source discrete signal data stream. 4. the distributed stream processing method that MFCC according to claim 3 extracts, it is characterized in that, described multi-source discrete signal data flow is carried out data flow partition and window operation, obtains parallel continuous sliding window, comprises as follows step: Performing a keyBy operation on the multi-source discrete signal data stream according to the sensor ID to obtain a keyed data stream; wherein, the keyBy operation is specifically sending the discrete signal data stream with the same sensor ID to a designated partition; A data stream windowing operation is performed on the keyed data stream in each partition to obtain parallel and continuous sliding windows corresponding to each sensor. 5. The distributed stream processing system extracted by MFCC is characterized in that, comprising: A data acquisition module, configured to acquire multi-source signal original data streams in parallel; wherein, the data type of the multi-source signal original data streams is String data; The data processing module is used to perform parallel flat mapping on the original data stream of the multi-source signal to obtain the data stream of the multi-source discrete signal; to perform data stream windowing operation on the data stream of the multi-source discrete signal to obtain continuous sliding windows ; Utilize the parallel window processing function to extract the Mel-frequency cepstral coefficient in the continuous sliding window, and obtain the Mel-frequency cepstral coefficient data flow; Utilize the parallel window processing function to extract the Mel-frequency cepstral coefficients in the continuous sliding window, and obtain the Mel-frequency cepstral coefficient data flow, comprising the following steps: storing data in each of the sliding windows in a corresponding double-precision array by using the parallel window processing function; calling the Mel-frequency cepstral coefficient extraction function for each of the double-precision arrays to obtain the Mel-frequency cepstral coefficient data stream; The Mel frequency cepstral coefficient extraction function includes a main function and a plurality of sub-functions, and a plurality of sub-functions are respectively a Mel filter bank function, a DCT function, a Fast Fourier Transform FFT function and a Hamming window function; The double-precision array is input into the main function, and the main function calculates the double-precision array by calling a plurality of the sub-functions to obtain Mel-frequency cepstral coefficients. 6. a storage medium, it is characterized in that, described storage medium is stored with computer program, when described computer program is carried out by the processor of computer, realizes the distributed method that MFCC extracts as any one of claims 1 to 4 stream processing method. 7. A kind of computer, it is characterized in that, comprise memory and processor, described memory is stored with computer program, when described computer program is carried out by described processor, realize MFCC as any one of claim 1 to 4 Extractive Distributed Stream Processing Approach.

Images