Industrial process oscillation detection method based on adaptive cyclic singular spectrum analysis

Abstract

The invention discloses an industrial process oscillation detection method based on adaptive cyclic singular spectrum analysis, and the method comprises the steps: firstly, carrying out the data normalization processing of a PV signal, secondly, determining the optimal window length according to the frequency information of the signal, decomposing the optimal window length into a series of RCs through employing an SCISSA method, and carrying out the analysis of the RCs. And based on the energy contribution degree and the Euclidean distance, obtaining a dominant component of a significant periodic characteristic in the PV signal, and finally calculating an energy concentration index (EI) for the signal to realize oscillation detection. According to the method, priori knowledge of an industrial process is not needed, the method has high robustness for non-stationary trend terms and noise in actually collected PV signals, the PV signals can be divided into different frequency components, oscillation RCs with practical significance can be extracted, and therefore high-precision oscillation detection is achieved.

Description

Technical Field

[0001] This invention relates to the field of oscillation detection algorithm technology, and in particular to an industrial process oscillation detection method based on adaptive cyclic singular spectrum analysis. Background Technology

[0002] Industrial processes are widely distributed in process industries, primarily those focused on basic raw materials such as petroleum, chemicals, steel, non-ferrous metals, and building materials. As the core of industrial production process control, the control system automatically adjusts parameters such as temperature, pressure, flow rate, and liquid level in real time to ensure stable and efficient operation. The proportional-integral-derivative (PID) controller, due to its simple structure, stable operation, robustness, and fast response, has become the most widely used controller in industrial process control, covering over 85% of control loops. However, with the passage of time, changes in the coupling characteristics of complex production processes and equipment lead to a gradual degradation of controller performance. Oscillation is one of the main manifestations of control loop performance degradation. This instability not only reduces product quality but also increases raw material and energy consumption, accelerates equipment aging, and even threatens production safety. Therefore, timely detection of oscillation and making maintenance decisions are crucial for maintaining production process stability and improving economic efficiency.

[0003] Existing oscillation detection methods have limitations in practical applications. Most methods require process data to possess strict preconditions such as stationarity, noise-free operation, and linear time-invariance. However, in actual production processes, due to complex factors such as external disturbances, sensor failures, and loop coupling effects, process data often fails to meet these prior conditions, leading to a decrease in the reliability of oscillation detection. Existing methods fail to effectively handle noise and non-stationary trend terms in the original signal, and directly performing oscillation detection on signals containing interference easily results in false positives and false negatives. Furthermore, existing methods lack adaptability to different operating conditions and cannot adapt to the diversity and complexity of industrial control processes.

[0004] Therefore, there is an urgent need to develop an industrial process oscillation detection method based on adaptive cyclic singular spectrum analysis. Summary of the Invention

[0005] This invention provides a method for detecting oscillations in industrial processes based on adaptive cyclic singular spectrum analysis, in order to solve the aforementioned problems existing in the prior art.

[0006] To achieve the above objectives, the present invention provides the following technical solution: An industrial process oscillation detection method based on adaptive cyclic singular spectrum analysis includes: S1: Collect process variable signals from the industrial process control system, normalize the process variable signals, and generate normalized signals; S2: Perform spectral analysis on the normalized signal, and adaptively determine the decomposition window length based on the main peak frequency and the corresponding power spectral density value, according to the normalized signal length and adjustment parameters; S3: Based on the decomposition window length, calculate the second-order moment sequence of the samples, construct the cyclic matrix and perform eigenvalue decomposition, and generate the reconstructed component sequence through the operation of the eigenvector and the normalized signal. S4: Calculate the energy contribution based on the eigenvalues ​​corresponding to the reconstructed components, determine the reference reconstructed component, calculate the Euclidean distance between other reconstructed components and the reference reconstructed component, select the reconstructed components that meet the distance conditions and superimpose them to generate the reconstructed signal; S5: Perform Discrete Fourier Transform on the reconstructed signal, calculate the ratio of maximum power to average power in the positive frequency range as the energy concentration index, and determine the oscillation state of the process variable signal based on the comparison result of the energy concentration index and the threshold.

[0007] Furthermore, step S2 includes: S21: Perform spectral analysis on the normalized signal to obtain the frequency components and the corresponding power spectral density distribution; S22: Select a preset number of frequency components with the largest power values ​​from the power spectral density distribution as the main peak frequencies, and obtain the power spectral density values ​​corresponding to each main peak frequency; S23: Based on each main peak frequency, the corresponding power spectral density value, the normalized signal length, the first adjustment parameter, and the second adjustment parameter, the decomposition window length is determined by the adaptive window length calculation formula.

[0008] Furthermore, step S3 includes: S31: Calculate the sample second moment sequence based on the normalized signal and the optimal decomposition window length, according to the time delay sequence; S32: Based on the sample second-order moment sequence and the optimal decomposition window length, calculate the elements of the first row of the cyclic matrix using a weighted average method to construct the cyclic matrix; S33: Perform eigenvalue decomposition on the circular matrix to obtain the sequence of eigenvalues ​​and the corresponding sequence of eigenvectors; S34: Perform convolution operation on each feature vector and the normalized signal to generate reconstructed components corresponding to each feature vector, forming a set of reconstructed components.

[0009] Furthermore, step S4 includes: S41: Normalize each eigenvalue in the eigenvalue sequence and take its logarithm, calculate the energy contribution of each reconstructed component, and select the reconstructed component with the largest energy contribution as the benchmark reconstructed component. S42: Calculate the Euclidean distance in the time domain between each reconstructed component and the reference reconstructed component in the reconstructed component sequence; S43: Select the reconstructed components whose Euclidean distance meets the preset conditions as relevant reconstructed components; S44: The reference reconstruction component is superimposed with all relevant reconstruction components in the time domain to generate a reconstruction signal.

[0010] Furthermore, step S5 includes: S51: Perform a discrete Fourier transform on the reconstructed signal to generate a frequency domain complex sequence; S52: Calculate the power spectral density value of each frequency component in the positive frequency interval of the complex sequence in the frequency domain, obtain the maximum power spectral density value, and calculate the average value of all power spectral density values ​​in the positive frequency interval as the average power spectral density value. S53: Calculate the ratio of the maximum power spectral density value to the average power spectral density value to generate an energy concentration index; S54: Compare the energy concentration index with a preset threshold. If the energy concentration index is greater than the preset threshold, it is determined that the process variable signal has an oscillation phenomenon; otherwise, it is determined that there is no oscillation phenomenon.

[0011] Furthermore, in step S23, the optimal decomposition window length... Calculated using the following formula: in, It is the length of the normalized signal; The sampling frequency of the normalized signal; This indicates that the frontier signal was found through spectral analysis of the normalized signal. The frequencies of the main peaks, m is the preset number of main peak frequencies; β is the second adjustment parameter, and α is the first adjustment parameter. For the j-th dominant peak frequency, Let j be the power spectral density value corresponding to the j-th dominant peak frequency. This represents the total power of the normalized signal.

[0012] Furthermore, the first adjustment parameter α is 6, and the second adjustment parameter β is 0.6.

[0013] Furthermore, in step S31, the sample second-order moment sequence is calculated using the following formula: Where L1 is the length of the normalized signal, The sampled value of the normalized signal at time t. The sampled value of the normalized signal at time t+n The optimal decomposition window length; In step S32, the elements of the first row of the circular matrix are calculated using the following formula: in, and is an element in the sample second moment sequence.

[0014] Furthermore, in step S41, the energy contribution is calculated using the following formula: in, The eigenvalue corresponding to the k-th reconstructed component. The optimal decomposition window length; In step S42, the Euclidean distance is calculated using the following formula: in, Let i be the value of the i-th reconstructed component at time t. L1 is the value of the reference reconstructed component at time t, and L1 is the length of the normalized signal.

[0015] Furthermore, in step S53, the energy concentration index is calculated using the following formula: Where L1 is the length of the reconstructed signal, Let n be the power spectral density value of the nth frequency component. This represents the maximum power spectral density value within the positive frequency range; the summation operation is performed within the positive frequency range. The average power expressed in this frequency range.

[0016] Compared with the prior art, the present invention has the following advantages: This invention proposes a method for detecting oscillations in industrial processes based on Self-tuning Circulant Singular Spectrum Analysis (SCISSA). First, SCISSA is used to pre-isolate trend terms and noise in industrial process data, significantly improving the quality of the oscillation signal to be detected. Second, an oscillation behavior detection index based on energy distribution is proposed. Through this "preprocessing + detection" strategy, the detection accuracy of industrial process oscillation behavior is effectively improved.

[0017] Other features and advantages of the invention will be set forth in the description which follows, and will be apparent in part from the description, or may be learned by practicing the invention.

[0018] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0019] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart of an industrial process oscillation detection method based on adaptive cyclic singular spectrum analysis, as described in an embodiment of the present invention. Figure 2 This is a system flowchart in an embodiment of the present invention; Figure 3 This is a flowchart of the SCISSA process in an embodiment of the present invention; Figure 4 This is a flowchart illustrating the calculation of oscillation detection indicators in an embodiment of the present invention; Figure 5 This is a waveform diagram of a portion of the RCs obtained by SCISSA processing the simulated signal in an embodiment of the present invention; Figure 6 This is a diagram showing the result of SCISSA processing of simulated signals according to an embodiment of the present invention; Figure 7 This is a diagram showing the result of SCISSA processing real industrial signals in an embodiment of the present invention. Detailed Implementation

[0020] The preferred embodiments of the present invention will be described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0021] The embodiments of the present invention provide, as follows Figure 1 As shown, an industrial process oscillation detection method based on adaptive cyclic singular spectrum analysis includes: S1: Collect process variable signals from the industrial process control system, normalize the process variable signals, and generate normalized signals; S2: Perform spectral analysis on the normalized signal, and adaptively determine the decomposition window length based on the main peak frequency and the corresponding power spectral density value, according to the normalized signal length and adjustment parameters; S3: Based on the decomposition window length, calculate the second-order moment sequence of the samples, construct the cyclic matrix and perform eigenvalue decomposition, and generate the reconstructed component sequence through the operation of the eigenvector and the normalized signal. S4: Calculate the energy contribution based on the eigenvalues ​​corresponding to the reconstructed components, determine the reference reconstructed component, calculate the Euclidean distance between other reconstructed components and the reference reconstructed component, select the reconstructed components that meet the distance conditions and superimpose them to generate the reconstructed signal; S5: Perform Discrete Fourier Transform on the reconstructed signal, calculate the ratio of maximum power to average power in the positive frequency range as the energy concentration index, and determine the oscillation state of the process variable signal based on the comparison result of the energy concentration index and the threshold.

[0022] The following is a detailed description with reference to specific embodiments.

[0023] like Figure 2 As shown, this embodiment provides a method for detecting oscillations in industrial processes based on adaptive cyclic singular spectrum analysis, including the following steps: Acquire a set of PV signals from the industrial process control loop; The acquired PV signals are subjected to data normalization processing; Find the optimal window based on the frequency domain information of the PV signal. ; Applying SCISSA yields a set of RCs; Based on energy contribution Find the most significant signal in RCs Calculate the sum of all RCs European distance Find all The RCs, when superimposed, become the preprocessed signal. ; calculate Energy Concentration Index It also outputs the oscillation detection results.

[0024] Specifically, it includes: S1: Collect process variable signals from the industrial process control system, normalize the process variable signals, and generate normalized signals; When the control loop exhibits oscillatory behavior, its process variable PV will fluctuate accordingly. Based on this characteristic, to identify the oscillation phenomenon in the control loop, it is first necessary to collect PV signals from the industrial process control system. Simultaneously, to eliminate the influence of signal scale differences on subsequent analysis, the PV signals are normalized to obtain the normalized signal Y(t).

[0025] S2: Perform spectral analysis on the normalized signal. Based on the main peak frequency and the corresponding power spectral density value, adaptively determine the decomposition window length according to the normalized signal length and adjustment parameters. This step uses Adaptive Cyclic Singular Spectrum Analysis (SCISSA) to extract significant periodic components from the PV signal, aiming to eliminate the influence of noise, non-stationary trend terms, and other factors on oscillation detection. The first step is to determine the optimal decomposition window length based on the frequency domain characteristics of the normalized signal Y(t). This parameter primarily reflects the signal periodicity information; improper selection can lead to over- or under-decomposition of the signal.

[0026] Specifically, it includes the following sub-steps: S21: Perform spectral analysis on the normalized signal to obtain the frequency components and the corresponding power spectral density distribution; Spectral analysis is performed on the normalized signal Y(t) to obtain each frequency component and its corresponding power spectral density distribution P(v).

[0027] S22: Select a preset number of frequency components with the largest power values ​​from the power spectral density distribution as the main peak frequencies, and obtain the power spectral density values ​​corresponding to each main peak frequency; Extract the m frequency components with the largest power spectral density values ​​from the power spectral density distribution, and use these frequency components as the main peak frequencies. (j=1,…,m), where m is typically set to 3. Simultaneously, acquire each main peak frequency. Corresponding power spectral density value .

[0028] S23: Based on each main peak frequency, the corresponding power spectral density value, the normalized signal length, the first adjustment parameter, and the second adjustment parameter, the decomposition window length is determined by the adaptive window length calculation formula. The optimal decomposition window length is determined according to the following formula. : in, yes Signal length; For normalized signal The sampling frequency; Indicates the normalized signal Spectral analysis was used to find the front The frequency of the main peaks, It is usually set to 3; The power spectral density value corresponding to the j-th dominant peak frequency is... Power at the location; These are two parameters that can be adjusted according to the actual signal; the default setting is... , .

[0029] S3: Based on the decomposition window length, calculate the second-order moment sequence of the samples, construct the cyclic matrix and perform eigenvalue decomposition, and generate the reconstructed component sequence through the operation of the eigenvector and the normalized signal. Complete the SCISSA decomposition and obtain the optimal window length. Afterwards, as Figure 3 As shown, the normalized signal Y(t) is decomposed into a series of reconstructed components RCs through cyclic matrix construction and eigenvalue decomposition.

[0030] Specifically, it includes the following sub-steps: S31: Calculate the sample second-order moment sequence based on the normalized signal and the optimal decomposition window length, according to the time delay sequence. Using the sample second moments of the normalized signal Y(t) It can be calculated using the following formula: in, The length of the normalized signal, The sampled value of the normalized signal at time t. The sampled value of the normalized signal at time t+n This is the optimal decomposition window length.

[0031] S32: Based on the sample second-order moment sequence and the optimal decomposition window length, calculate the elements of the first row of the cyclic matrix using a weighted average method, and construct the cyclic matrix. Calculate the circular matrix The first row of elements : in, and Let be the elements in the sample second-order moment sequence. Based on the properties of a circular matrix, the elements in the first row... Construct a complete cyclic matrix .

[0032] S33: Perform eigenvalue decomposition on the circular matrix to obtain the sequence of eigenvalues ​​and the corresponding sequence of eigenvectors; For circular matrices Perform eigenvalue decomposition to obtain the eigenvalue sequence. (k=1,…, ) and the corresponding feature vector sequence Each feature vector Corresponding to a discrete frequency , representing a sinusoidal component with a fixed frequency. Eigenvalues Calculated using the following formula: in, Represents a circular matrix elements, Represents discrete frequency The power spectral density at that location.

[0033] S34: Perform convolution operation between each feature vector and the normalized signal to generate reconstructed components corresponding to each feature vector, forming a set of reconstructed components; Based on the eigenvector sequence, the normalized signal Y(t) is decomposed, and the original signal can finally be decomposed into a set of RCs: in, Let i be the i-th reconstructed component.

[0034] S4: Calculate the energy contribution based on the eigenvalues ​​corresponding to the reconstructed components, determine the reference reconstructed component, calculate the Euclidean distance between other reconstructed components and the reference reconstructed component, select reconstructed components that meet the distance conditions and superimpose them to generate a reconstructed signal; this step uses the energy contribution and Euclidean distance to optimize the selection strategy of RCs and realize the extraction of strong periodic signals.

[0035] Specifically, it includes the following sub-steps: S41: Normalize each eigenvalue in the eigenvalue sequence and take its logarithm. Calculate the energy contribution of each reconstructed component and select the reconstructed component with the largest energy contribution as the benchmark reconstructed component. Utilizing energy contribution Find the signal with the highest energy, i.e., the reconstruction component with the largest energy contribution, and use it as the benchmark reconstruction component. .

[0036] S42: Calculate the Euclidean distance in the time domain between each reconstructed component in the reconstructed component sequence and the reference reconstructed component. Calculate all reconstructed components RCs and the baseline reconstructed component Euclidean distance The calculation formula is as follows: in, Let i be the value of the i-th reconstructed component at time t. The value of the reference reconstructed component at time t, The length of the normalized signal.

[0037] S43: Select the reconstructed components whose Euclidean distance meets the preset conditions as relevant reconstructed components; When the Euclidean distance between two reconstructed components When the distance is greater than 0.9, the components can be considered to belong to the same signal component. Reconstructed components with a Euclidean distance greater than 0.9 are selected as relevant reconstructed components.

[0038] S44: Time-domain superposition of the reference reconstructed component and all relevant reconstructed components to generate a reconstructed signal; Find all components reconstructed from the baseline Euclidean distance The reconstructed components RCs are then superimposed in the time domain with the reference reconstructed components to obtain the reconstructed signal. .

[0039] S5: Perform a Discrete Fourier Transform on the reconstructed signal, calculate the ratio of maximum power to average power within the positive frequency range as an energy concentration index, and determine the oscillation state of the process variable signal based on the comparison result of the energy concentration index and the threshold; this step calculates the reconstructed signal. Energy concentration index EI, such as Figure 3 As shown, oscillation detection is achieved using the SCISSA algorithm.

[0040] Specifically, it includes the following sub-steps: S51: Perform a discrete Fourier transform on the reconstructed signal to generate a frequency domain complex sequence; For reconstructed signals Perform a Discrete Fourier Transform (DFT) to obtain a complex sequence X[n] in the frequency domain, where n is the frequency index.

[0041] S52: Calculate the power spectral density value of each frequency component in the positive frequency interval of the complex sequence in the frequency domain, obtain the maximum power spectral density value, and calculate the average value of all power spectral density values ​​in the positive frequency interval as the average power spectral density value. Calculate the positive frequency range ( Power spectral density values ​​of each frequency component within ) Obtain the maximum power spectral density value Calculate the average value of all power spectral density values ​​within this frequency range. As the average power spectral density value.

[0042] S53: Calculate the ratio of the maximum power spectral density value to the average power spectral density value to generate an energy concentration index, such as... Figure 4 As shown, calculate the oscillation detection index; The energy concentration index EI is calculated as follows: in, For reconstructing signals The signal length, Let n be the power spectral density value of the nth frequency component. This indicates the maximum power value within the positive frequency range. This represents the average power within that frequency range.

[0043] S54: Compare the energy concentration index with a preset threshold. If the energy concentration index is greater than the preset threshold, determine that the process variable signal has an oscillation phenomenon; otherwise, determine that there is no oscillation phenomenon. If reconstructed signal If EI>0.8, then the process variable signal Y(t) is considered to have oscillation, meaning the control loop is in an abnormal operating condition; otherwise, no oscillation is considered to exist.

[0044] To verify the effectiveness of this invention, this embodiment first analyzes a simulated signal. The simulated signal is constructed as follows: in: in It is Gaussian white noise, that is The data length is 1024, and the sampling frequency is 1000Hz.

[0045] First, y(t) is normalized to obtain the normalized signal Y(t). Second, the normalized signal Y(t) is subjected to spectral analysis to extract the first three main peak frequencies and their corresponding power spectral density values. The optimal decomposition window length is then calculated according to formula (1). .

[0046] Next, according to steps S31-S32, a cyclic matrix is ​​constructed using the sample second-order moment sequence. Following step S33, perform eigenvalue decomposition on the circulant matrix to obtain eigenvalues. (k=1,…,352) and a sequence of eigenvectors. According to step S34, the original signal is decomposed into a set of reconstructed components RCs, some of which (the first 12) are as follows: Figure 5 As shown.

[0047] According to step S41, the energy contribution index is used. The reconstructed component with the largest energy contribution among the reconstructed components RCs is determined as the benchmark reconstructed component. Following steps S42-S44, relevant reconstructed components are selected based on Euclidean distance and then superimposed to obtain the reconstructed signal. The result is as follows Figure 6 As shown in the figure, the proposed method performs excellently in extracting oscillating signals, successfully eliminating non-stationary trend terms and noise interference in the simulated signal, and completely extracting the characteristic information of the oscillating signal.

[0048] According to steps S51-S53, the reconstructed signal obtained after SCISSA preprocessing is processed... Perform a Discrete Fourier Transform and calculate the corresponding energy concentration index EI, as shown in Table 1. From the table, it can be observed that the reconstructed signal... The energy concentration index EI = 0.9854 > 0.8, which meets the oscillation judgment condition. According to step S54, the simulation signal is determined to have oscillation. The proposed oscillation detection method successfully identified the oscillation phenomenon in the simulation case.

[0049] Table 1. Oscillation detection results of the simulated signal using the method of the present invention. This embodiment uses the process variable PV from Chem.loop35 in the industrial dataset isdb10 for analysis. Following step S1 of the method of this invention, the acquired PV signal is first normalized. Then, following step S2, the normalized signal undergoes spectral analysis to adaptively determine the optimal decomposition window length. Next, following step S3, the normalized signal is decomposed into a series of reconstructed components RCs using the SCISSA method.

[0050] Following step S4, relevant reconstructed components are selected and superimposed using energy contribution and Euclidean distance to obtain the reconstructed signal. ,like Figure 7 As shown. From Figure 7 It can be observed that SCISSA completely preserves the oscillation characteristics of PV data in the reconstructed signal. In China, effectively and accurately extracting oscillation features from PV data is crucial for achieving precise oscillation detection.

[0051] Finally, following step S5, the reconstructed signal... Discrete Fourier Transform was performed to calculate the energy concentration index EI, and the results are shown in Table 2. According to the oscillation detection index provided in Table 2, the reconstructed signal's EI = 0.8956 > 0.8, therefore it is identified as an oscillation. The method proposed in this invention successfully detected oscillation phenomena in real industrial signals.

[0052] Table 2. Oscillation detection results of the method of the present invention on real industrial signals. The SCISSA algorithm proposed in this invention utilizes the correspondence between the eigenvalues ​​of the constructed cyclic matrix and the discrete frequencies of the original signal to classify signals of different frequencies into RCs, effectively avoiding the mode aliasing problem caused by traditional decomposition algorithms. The proposed SCISSA algorithm first determines the optimal window length L_b by comprehensively considering factors such as signal frequency information and power, avoiding over- or under-decomposition of the signal and successfully classifying the signal into different RCs. Then, it optimizes the RCs selection strategy, using energy contribution and Euclidean distance between signals to successfully restore periodic features masked by noise and non-stationary trend terms. The EI proposed in this invention relies solely on the signal's DFT results to identify potential frequency bands of periodic oscillations, calculating the ratio of the maximum single-frequency power to the average power of the same frequency band, thereby identifying strong periodic oscillation signals in the signal. This method requires no prior knowledge, has simple steps, and possesses high adaptability and stability, offering significant advantages over traditional methods.

[0053] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from the spirit and scope of this invention.

Claims

1. A method for detecting oscillations in an industrial process based on adaptive cyclic singular spectrum analysis, characterized in that, Comprise: S1: collect the process variable signal of the industrial process control system, normalize the process variable signal, and generate a normalized signal; S2: perform spectral analysis on the normalized signal, and based on the main peak frequency and the corresponding power spectral density value, determine the decomposition window length adaptively according to the normalized signal length and the adjustment parameter; S3: based on the decomposition window length, calculate the sample second moment sequence, construct the circulant matrix and perform eigenvalue decomposition, and generate the reconstructed component sequence through the operation of the normalized signal and the eigenvector; S4: calculate the energy contribution degree of the reconstructed component based on the corresponding eigenvalue, determine the reference reconstructed component, calculate the Euclidean distance between the reference reconstructed component and other reconstructed components in the time domain, select the reconstructed components that meet the distance condition and superimpose them to generate the reconstructed signal; S5: perform discrete Fourier transform on the reconstructed signal, calculate the ratio of the maximum power to the average power in the positive frequency interval as the energy concentration index, and determine the oscillation state of the process variable signal according to the comparison result of the energy concentration index and the threshold value.

2. The method for industrial process oscillation detection based on adaptive cyclic singular spectrum analysis according to claim 1, characterized in that, The step S2 comprises: S21: perform spectral analysis on the normalized signal to obtain the frequency component and the corresponding power spectral density distribution; S22: select the preset number of frequency components with the maximum power value from the power spectral density distribution as the main peak frequency, and obtain the power spectral density value corresponding to each main peak frequency; S23: based on each main peak frequency, the corresponding power spectral density value, the normalized signal length, the first adjustment parameter and the second adjustment parameter, determine the decomposition window length through the adaptive window length calculation formula.

3. The method for industrial process oscillation detection based on adaptive cyclic singular spectrum analysis according to claim 1, characterized in that, The step S3 comprises: S31: based on the normalized signal and the optimal decomposition window length, calculate the sample second moment sequence according to the time delay sequence; S32: based on the sample second moment sequence and the optimal decomposition window length, calculate the first row elements of the circulant matrix through weighted average method to construct the circulant matrix; S33: perform eigenvalue decomposition on the circulant matrix to obtain the eigenvalue sequence and the corresponding eigenvector sequence; S34: perform convolution operation on each eigenvector and the normalized signal to generate the reconstructed component corresponding to each eigenvector, and form the reconstructed component set.

4. The method for industrial process oscillation detection based on adaptive cyclic singular spectrum analysis according to claim 1, characterized in that, The step S4 comprises: S41: normalize each eigenvalue in the eigenvalue sequence and take the logarithm, calculate the energy contribution degree of each reconstructed component, and select the reconstructed component with the maximum energy contribution degree as the reference reconstructed component; S42: calculate the Euclidean distance between each reconstructed component in the reconstructed component sequence and the reference reconstructed component in the time domain; S43: select the reconstructed components with the Euclidean distance meeting the preset condition as the related reconstructed components; S44: superimpose the reference reconstructed component and all related reconstructed components in the time domain to generate the reconstructed signal.

5. The adaptive cyclic singular spectrum analysis based industrial process oscillation detection method according to claim 1, wherein, The step S5 comprises: S51: perform discrete Fourier transform on the reconstructed signal to generate a frequency domain complex sequence; S52: calculate the power spectral density value of each frequency component in the positive frequency interval in the frequency domain complex sequence, obtain the maximum power spectral density value, and calculate the average value of all power spectral density values in the positive frequency interval as the average power spectral density value; S53: calculate the ratio of the maximum power spectral density value to the average power spectral density value to generate the energy concentration index; S54: compare the energy concentration index with a preset threshold value, when the energy concentration index is greater than the preset threshold value, it is determined that the process variable signal has oscillation phenomenon, otherwise it is determined that there is no oscillation phenomenon.

6. The method for industrial process oscillation detection based on adaptive cyclic singular spectrum analysis according to claim 2, characterized in that, In step S23, the optimal decomposition window length is calculated by the following equation: ； in, It is the length of the normalized signal; The sampling frequency of the normalized signal; This indicates that the frontier signal was found through spectral analysis of the normalized signal. The frequencies of the main peaks, m is the preset number of main peak frequencies; β is the second adjustment parameter, and α is the first adjustment parameter. For the j-th dominant peak frequency, Let j be the power spectral density value corresponding to the j-th dominant peak frequency. This represents the total power of the normalized signal.

7. The adaptive cyclic singular spectrum analysis based industrial process oscillation detection method according to claim 6, characterized in that, The first adjustment parameter a is 6, and the second adjustment parameter b is 0.

6.

8. The adaptive cyclic singular spectrum analysis based industrial process oscillation detection method according to claim 3, wherein, In step S31, the sample second moment sequence is calculated by the following formula: ； wherein L1 is the length of the normalized signal, is the sampling value of the normalized signal at time t, is the sampling value of the normalized signal at time t+n, is the optimal decomposition window length; In step S32, the first row element of the circulant matrix is calculated by the following formula: ； wherein and are elements in the sample second moment sequence.

9. The adaptive cyclic singular spectrum analysis based industrial process oscillation detection method according to claim 4, wherein, In step S41, the energy contribution degree is calculated by the following formula: ； wherein, is the eigenvalue corresponding to the kth reconstructed component, is the optimal decomposition window length; In step S42, the Euclidean distance is calculated by the following formula: ； wherein is the value of the i-th reconstructed component at time t, is the value of the reference reconstructed component at time t, and L1is the length of the normalized signal.

10. The adaptive cyclic singular spectrum analysis based industrial process oscillation detection method according to claim 5, wherein, In step S53, the energy concentration index is calculated by the following formula: ； where L1 is the length of the reconstructed signal, is the power spectral density value of the nth frequency component, is the maximum power spectral density value within the positive frequency range, the summation operation being performed within the positive frequency range, is the average power present in the frequency range.

Images