Marine rocket erection state anomaly detection method based on sparse representation and adaptive filtering

Abstract

The invention discloses an offshore rocket erection state anomaly detection method based on sparse representation and adaptive filtering, and the method comprises the following steps: obtaining state data in a rocket erection process through multiple sensors, constructing a time sequence, setting a sliding window with change capability in each time period, and extracting sparse features for state prediction. A feedback adjustment signal is formed by calculating the difference between a predicted value and an observed value, and parameters in the state estimation process and the external disturbance modeling process are adjusted respectively. A double-path structure is adopted to independently estimate the state and disturbance, and a final estimation result is obtained through a fusion strategy. The system continuously updates parameters according to error conditions, continuous self-adaptive adjustment and anomaly recognition are achieved, and the recognition precision and the response capability of the anomaly trend are improved. The method is suitable for the recognition processing of the abnormal state in the erection process of the rocket in the marine environment, and has high dynamic adaptive capacity and time sequence anomaly detection capacity.

Description

Technical Field

[0001] The present invention relates to the technical field of aerospace measurement and control and intelligent signal processing, and in particular to a method for detecting abnormalities in the erection state of a sea-based rocket based on sparse representation and adaptive filtering. Background Art

[0002] As maritime space launch missions become increasingly commonplace, the rocket erection process in complex sea conditions has become a crucial factor impacting mission safety and success. Timely detection and accurate identification of rocket erection state anomalies are crucial for ensuring structural stability, avoiding abnormalities, and improving the controllability of the erection process. However, the uncertainties of wind and wave disturbances, high sensor measurement noise, and highly nonlinear structural response in maritime environments present significant challenges in practical application of traditional state detection methods.

[0003] In existing technologies, common state anomaly detection methods rely on fixed threshold judgment, single-channel data monitoring, or linear filter estimation. These methods are difficult to effectively capture weak anomalies or dynamic trends when faced with multi-source information fusion, dynamic changes in disturbances, and high-noise environments during the erection process. They have the following drawbacks:

[0004] 1. Static judgment mechanism lags: Traditional methods mostly use static thresholds or fixed judgment rules, which cannot cope with nonlinear offsets and small trend variations in the state evolution process and lack dynamic response capabilities.

[0005] 2. Weak multi-sensor information fusion capability: Existing methods find it difficult to achieve collaborative modeling and compensation of multi-channel state information, especially when the data is incomplete or some sensors fail, the state estimation accuracy drops significantly.

[0006] 3. Lack of adaptability of filtering structure: Traditional filtering methods do not distinguish between paths for processing disturbance and state common-mode signals, and cannot distinguish between state evolution and disturbance response, which can easily lead to misestimation or delayed response.

[0007] 4. Single parameter adjustment mechanism: Parameters in the filtering and anomaly detection process are mostly set manually, lacking an adaptive adjustment strategy based on real-time feedback, making it difficult to adapt to unstable systems under sea disturbances.

[0008] 5. Limited anomaly recognition accuracy: In the context of high interference and low signal-to-noise ratio, traditional methods have difficulty accurately extracting weak anomaly features from residuals, resulting in high missed or misjudgment rates.

[0009] Therefore, how to provide a method for detecting abnormalities in the erection state of a sea-based rocket based on sparse representation and adaptive filtering is an urgent problem that needs to be solved by those skilled in the art. Summary of the Invention

[0010] One purpose of the present invention is to propose a method for detecting abnormalities in the erection state of a sea-based rocket based on sparse representation and adaptive filtering. The present invention constructs a multi-channel dynamic sliding window mechanism, combines sparse coding to extract state features, and introduces a dual-path filtering structure of state prediction and disturbance compensation. The adaptive adjustment of the filter gain and modeling parameters is achieved through residual feedback regulation, which has the advantages of high recognition accuracy, strong dynamic adaptability and real-time anomaly detection.

[0011] According to an embodiment of the present invention, a method for detecting abnormalities in the erection state of a marine rocket based on sparse representation and adaptive filtering includes the following steps:

[0012] S1. Obtain the original state observation data collected by multiple state monitoring sensors during the erection process of the sea rocket, perform time alignment processing, and generate a multi-channel state observation time series sequence;

[0013] S2. Construct a dynamic sliding window mechanism based on the multi-channel state observation time series, set an adaptive window length for each dynamic sliding window, and generate a dynamic sparse representation dictionary;

[0014] S3. Sparsely encode the state observation data in the dynamic sliding window based on the dynamic sparse representation dictionary, extract the state sparse feature vector, and calculate the current state prediction value based on the state sparse feature vector;

[0015] S4, calculating the residual signal between the current state prediction value and the actual state observation value, and calculating the feedback adjustment coefficient based on the residual signal;

[0016] S5. Construct a dual-path filtering architecture including a state prediction path and a disturbance compensation path, wherein the state prediction path receives a state sparse feature vector for performing state estimation, and the disturbance compensation path receives a residual signal for performing disturbance estimation, and the outputs of the dual paths are fused to generate a final state estimation result;

[0017] S6. According to the feedback adjustment coefficient, the filter gain parameter used in state estimation and the disturbance modeling parameter used in disturbance estimation are adjusted in a coordinated manner to establish a parameter adaptive adjustment mechanism;

[0018] S7. Input the final state estimation result and the residual signal into an anomaly detection model. The anomaly detection model identifies state anomalies during the erection of the sea-based rocket based on a time series consistency analysis method and outputs an anomaly labeling result.

[0019] Optionally, the S2 specifically includes:

[0020] S21, performing unified time axis mapping on the multi-channel state observation time series to construct a multi-channel observation matrix covering the synchronous data of each channel;

[0021] S22. In the multi-channel observation matrix, set the initial position of the sliding window, and extract the state observation data of each channel in the corresponding time period as the data input segment corresponding to the current sliding window;

[0022] S23. For the data input segment corresponding to the current sliding window, calculate the state change rate of each channel in the time period, and construct a nonlinear window length adjustment function based on the average state change rate of all channels. Define the current window length L(t) as:

[0023]

[0024] Among them, L min is the minimum length of the sliding window, L max is the maximum length of the sliding window, v i (t) is the state change rate of the i-th channel in the current time period, and N is the total number of state monitoring channels;

[0025] S24. Calculate the sliding residual signal between the state prediction value and the actual state observation value in the previous sliding cycle of the current sliding window, and calculate the change amplitude Δε of the sliding residual signal. If the change amplitude Δε is greater than the set residual change threshold θ, move the starting position of the sliding window forward by a sliding step Δt, and update the sliding window position. If Δε is less than or equal to the residual change threshold θ, keep the starting position of the sliding window unchanged.

[0026] S25. Extract observation data segments within each triggered sliding window, and construct a set of dynamic sparse representation dictionaries corresponding to the time features of each sliding window.

[0027] Optionally, the S3 specifically includes:

[0028] S31, extracting state observation data segments within the time period corresponding to each sliding window, and constructing a data matrix to be encoded, wherein each column of the data matrix to be encoded corresponds to a time series sample of a state monitoring channel;

[0029] S32. Taking the data matrix to be encoded as input, combining it with the dynamic sparse representation dictionary corresponding to the sliding window, using the orthogonal matching pursuit algorithm to perform sparse encoding on each state observation data, solving the sparse coefficient vector α, satisfying the following expression: x≈D·α, where x is the state observation vector, D is the dynamic sparse representation dictionary, and α is the sparse coding coefficient vector, and extracting the state sparse feature vector;

[0030] S33. Input the sparse coding coefficient vector α into the state prediction structure. The state prediction structure is composed of two nonlinear mapping functions, including a gate activation function and a state projection function:

[0031]

[0032] in, is the current state prediction value, W g 、W p are the weight matrices of the gated path and the predicted path, b g 、b p is the corresponding bias term, σ(·) is the nonlinear activation function;

[0033] S34. Arrange the state prediction values ​​generated by the continuous sliding windows in chronological order to construct a state prediction sequence.

[0034] Optionally, the S4 specifically includes:

[0035] S41. Obtain the current state prediction value in the constructed state prediction sequence And the actual state observation value y at the corresponding moment t , as the current predicted value and the current observed value respectively;

[0036] S42. Predicted value based on current state and the actual state observation value y t The difference between the residual signal ε is calculated t ;

[0037] S43, based on the residual signal sequence {ε t-n ,…,ε t}, calculate the residual change rate Δε t , the residual change rate is the difference between the residual signal at the current moment and the residual signal at the previous moment;

[0038] S44, according to the residual signal ε t and the residual change rate Δε t , construct the feedback regulation function f(·), and calculate the feedback regulation coefficient γ t , the feedback adjustment function adopts the form of exponential weight combination, which is expressed as follows:

[0039] γ t =exp(-λ1|ε t |-λ2|Δε t |);

[0040] Among them, γ t is the feedback adjustment coefficient, λ1 and λ2 are the adjustment weight coefficients of the residual amplitude and change rate, respectively, which are used to control the response degree to the residual intensity and change trend.

[0041] Optionally, the S5 specifically includes:

[0042] S51. Receive the state sparse feature vector, use the state sparse feature vector as the input of the state prediction path, and use the prediction model in the Kalman filter structure to perform state estimation to generate a state estimation value:

[0043]

[0044] in, is the estimated value of the current state, is the state estimate at the previous moment, A is the state transfer matrix, B is the control input matrix, u t is an optional control input;

[0045] S52: Receive the residual signal, use the residual signal as the input of the disturbance compensation path, and use the exponentially weighted residual modeling method to perform disturbance estimation to generate a disturbance estimation value:

[0046]

[0047] Among them, d t is the current disturbance estimate, ε t-i is the residual signal before the i-th step, ω i Estimate the weight coefficient for the disturbance to meet the normalization condition;

[0048] S53. Perform weighted fusion on the state estimation value generated by the state prediction path and the disturbance estimation value generated by the disturbance compensation path to construct a final state estimation value. The fusion method adopts a weighted average model, which is expressed as:

[0049]

[0050] in, is the current final state estimate, and λ is the fusion weight coefficient, which balances the state estimation and disturbance estimation.

[0051] Optionally, the S6 specifically includes:

[0052] S61, receiving a feedback adjustment coefficient, and inputting the feedback adjustment coefficient into the state prediction path and the disturbance compensation path respectively, for linkage adjustment of the filter gain parameter in the state estimation path and the disturbance modeling parameter in the disturbance compensation path;

[0053] S62. In the state prediction path, dynamically update a filter gain parameter based on the current value of the feedback adjustment coefficient. The filter gain parameter is used in the state update step in the Kalman filter structure. When the feedback adjustment coefficient increases, the filter gain parameter is correspondingly increased to enhance the ability to track rapid state changes. When the feedback adjustment coefficient decreases, the filter gain parameter is correspondingly decreased to suppress invalid responses.

[0054] S63. In the disturbance compensation path, adjusting disturbance modeling parameters according to a change trend of the current feedback adjustment coefficient, wherein the disturbance modeling parameters include a time-weighted window length for disturbance estimation and a weighted distribution structure of historical residual signals. When the feedback adjustment coefficient is increased, the proportion of the current residual signal in the disturbance estimation is increased, and when the feedback adjustment coefficient is decreased, the influence of historical disturbance trend information is increased.

[0055] S64. Within the current sliding window period, respectively calculate the state estimation error between the state estimation value output by the state prediction path and the actual state observation value, and the disturbance estimation error between the disturbance estimation value output by the disturbance compensation path and the actual state observation value, construct a joint error evaluation index, and establish a parameter adjustment criterion with the goal of minimizing the joint error;

[0056] S65. Based on the output result of the joint error evaluation index, dynamically adjust the adjustment step size and update direction of the filter gain parameter used in the state prediction path and the disturbance modeling parameter used in the disturbance compensation path, and establish a linked coupled dual-factor parameter adaptive adjustment mechanism;

[0057] S66. Apply the filter gain parameters and disturbance modeling parameters updated by the joint error drive to the execution process of the state prediction path and the disturbance compensation path at the next moment, forming a continuous parameter self-optimization closed-loop system driven by the feedback adjustment coefficient.

[0058] Optional, S65 specifically includes:

[0059] S651. Generate a joint error evaluation index based on the calculated state estimation error and disturbance estimation error, where the joint error evaluation index is composed of a weighted combination of the state estimation error and the disturbance estimation error.

[0060] S652: Compare the joint error evaluation index with a preset joint error threshold, and determine whether to initiate a parameter adjustment process based on the comparison result. If the joint error evaluation index is greater than the preset joint error threshold, perform parameter adjustment; if not, maintain the current parameter settings.

[0061] S653: In parameter adjustment, the change value of the state estimation error is called as an adjustment control variable and input into the state prediction path to adjust the filter gain parameter. The adjustment operation sets the adjustment step size and direction according to the error change value to update the filter gain parameter;

[0062] S654: Synchronously call the change value of the disturbance estimation error and transmit it to the disturbance compensation path as a control input to adjust the disturbance modeling parameters. The disturbance modeling parameter adjustment strategy includes adjusting the time-weighted distribution of the residual signal and the estimation window span;

[0063] S655: Jointly analyze the error change value of the state prediction path and the error change value of the disturbance compensation path to calculate a joint adjustment weighting factor. The joint adjustment weighting factor is used to allocate adjustment resources between the two paths.

[0064] S656: Output the updated filter gain parameters and disturbance modeling parameters respectively and pass them to S66 to build a continuous dual-path parameter linkage update chain.

[0065] The beneficial effects of the present invention are:

[0066] (1) This invention introduces a dynamic sliding window mechanism and a sparse representation method, combined with a nonlinear gated prediction structure and an orthogonal matching pursuit algorithm, to achieve fine-grained extraction and dynamic modeling of the erection state characteristics of sea-based rockets, significantly improving the state recognition capability under high noise and strong disturbance backgrounds. Based on the residual feedback adjustment mechanism, a linkage update process of the filter gain parameters and the disturbance modeling parameters is constructed, and a dual-factor coupling adjustment mechanism is established between the state prediction path and the disturbance compensation path, so that the system has real-time parameter optimization and structural self-adaptation capabilities, thereby improving the stability of state estimation and the accuracy of anomaly detection.

[0067] (2) The present invention utilizes a parameter adjustment criterion driven by a joint error evaluation index to achieve dynamic coordinated control of the filtering path and the perturbation path. Furthermore, by constructing a residual activation and sliding window adjustment strategy, the algorithm is able to continuously track minor anomaly trends, effectively reducing the misjudgment rate caused by sensor drift or incomplete data. Combined with the temporal consistency determination method of the anomaly detection model, this method can achieve multi-dimensional anomaly perception throughout the entire rocket erection process, and can pre-label and grade potential fault trends.

[0068] (3) The present invention realizes structured data reception, feature extraction, prediction estimation, disturbance compensation, adjustment optimization and abnormality judgment closed-loop control in the entire erection process state recognition and abnormality monitoring process. It does not rely on static thresholds or single model responses, and has cross-model self-update, self-adjustment and weak signal recognition capabilities. It is suitable for highly robust state perception tasks under complex sea conditions. BRIEF DESCRIPTION OF THE DRAWINGS

[0069] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings:

[0070] Figure 1 This is a flow chart of the method for detecting abnormalities in the erection state of a marine rocket based on sparse representation and adaptive filtering proposed by the present invention. DETAILED DESCRIPTION

[0071] The present invention will now be described in further detail with reference to the accompanying drawings, which are simplified schematic diagrams that illustrate the basic structure of the present invention in a schematic manner.

[0072] refer to Figure 1 The method for detecting abnormalities in the erection state of a sea-based rocket based on sparse representation and adaptive filtering includes the following steps:

[0073] S1. Obtain the original state observation data collected by multiple state monitoring sensors during the erection process of the sea rocket, perform time alignment processing, and generate a multi-channel state observation time series sequence;

[0074] This implementation method deploys multiple state monitoring sensors on the offshore rocket erection platform to collect real-time multi-dimensional raw state observation data on the rocket's structure, attitude, environmental disturbances, and other aspects. Due to differences in sampling frequency and timestamp offsets among sensors, interpolation resampling and a unified time base alignment algorithm are used to time-align the raw data, construct an observation matrix on a unified time axis, and ensure that the data from each channel are comparable at the same time point. The resulting multi-channel state observation time series provides basic data support for subsequent sliding window partitioning, sparse coding, and anomaly detection, achieving effective conversion of multi-source asynchronous data to structured time series input, and improving the consistency and accuracy of state feature extraction.

[0075] S2. Construct a dynamic sliding window mechanism based on the multi-channel state observation time series, set an adaptive window length for each dynamic sliding window, and generate a dynamic sparse representation dictionary;

[0076] S3. Sparsely encode the state observation data in the dynamic sliding window based on the dynamic sparse representation dictionary, extract the state sparse feature vector, and calculate the current state prediction value based on the state sparse feature vector;

[0077] S4, calculating the residual signal between the current state prediction value and the actual state observation value, and calculating the feedback adjustment coefficient based on the residual signal;

[0078] S5. Construct a dual-path filtering architecture including a state prediction path and a disturbance compensation path, wherein the state prediction path receives a state sparse feature vector for performing state estimation, and the disturbance compensation path receives a residual signal for performing disturbance estimation, and the outputs of the dual paths are fused to generate a final state estimation result;

[0079] S6. According to the feedback adjustment coefficient, the filter gain parameter used in state estimation and the disturbance modeling parameter used in disturbance estimation are adjusted in a coordinated manner to establish a parameter adaptive adjustment mechanism;

[0080] S7. Input the final state estimation result and the residual signal into an anomaly detection model. The anomaly detection model identifies state anomalies during the erection of the sea-based rocket based on a time series consistency analysis method and outputs an anomaly labeling result.

[0081] This embodiment uses the final state estimation result and the residual signal as input to construct an anomaly detection model. By jointly analyzing the time evolution consistency between the state estimation sequence and the residual signal sequence, the continuity, trend deviation and residual mutation characteristics within the time window are extracted, and the dynamic threshold perception mechanism and the sliding correlation matching method are used to identify abnormal change patterns in the erection process. When there is a significant inconsistency between the predicted state and the observed data, and the inconsistency persists in multiple time steps, the model automatically determines it as a potential anomaly and outputs the corresponding anomaly labeling result. This method realizes multi-dimensional perception of weak variation trends and sudden anomalies, and improves the accuracy and stability of identifying abnormalities in the rocket erection state.

[0082] In this embodiment, S2 specifically includes:

[0083] S21, performing unified time axis mapping on the multi-channel state observation time series to construct a multi-channel observation matrix covering the synchronous data of each channel;

[0084] S22. In the multi-channel observation matrix, set the initial position of the sliding window, and extract the state observation data of each channel in the corresponding time period as the data input segment corresponding to the current sliding window;

[0085] S23. For the data input segment corresponding to the current sliding window, calculate the state change rate of each channel in the time period, and construct a nonlinear window length adjustment function based on the average state change rate of all channels. Define the current window length L(t) as:

[0086]

[0087] Among them, L min is the minimum length of the sliding window, L max is the maximum length of the sliding window, v i (t) is the state change rate of the i-th channel in the current time period, and N is the total number of state monitoring channels;

[0088] The formula uses the hyperbolic tangent function to construct a continuously differentiable nonlinear response curve. This ensures that when the rate of state change is low, the window length remains at a baseline level. When the rate of change increases significantly, the window length gradually shrinks within a limited range, thereby enhancing the system's sensitivity to sudden changes in state. This design effectively balances the stability and response speed of feature extraction, adaptively adjusting the sliding window granularity under varying state fluctuations, and improving the timeliness and accuracy of sparse feature representation.

[0089] S24. Calculate the sliding residual signal between the state prediction value and the actual state observation value in the previous sliding cycle of the current sliding window, and calculate the change amplitude Δε of the sliding residual signal. If the change amplitude Δε is greater than the set residual change threshold θ, move the starting position of the sliding window forward by a sliding step Δt, and update the sliding window position. If Δε is less than or equal to the residual change threshold θ, keep the starting position of the sliding window unchanged.

[0090] S25. Extract observation data segments within each triggered sliding window, and construct a set of dynamic sparse representation dictionaries corresponding to the time features of each sliding window.

[0091] This embodiment constructs a synchronous observation matrix by mapping the multi-channel state observation time series to a unified time axis, and sets a sliding window on the matrix to calculate the state change rate of each channel for the data segment in each window. A nonlinear function is used to adaptively adjust the window length, and the amplitude of the sliding residual signal change is introduced as a dynamic trigger condition for window sliding, making the sliding mechanism responsive and flexible. After the window sliding is triggered, the corresponding observation data is extracted and a dynamic sparse representation dictionary matching the time characteristics is constructed to achieve efficient expression and modeling of local state changes. This embodiment improves the method's temporal adaptability and feature expression accuracy in the process of continuous modeling of the erection state by dynamically adjusting the sliding window structure and the sparse dictionary update mechanism, laying a solid data processing foundation for subsequent state prediction and anomaly detection.

[0092] In this embodiment, S3 specifically includes:

[0093] S31, extracting state observation data segments within the time period corresponding to each sliding window, and constructing a data matrix to be encoded, wherein each column of the data matrix to be encoded corresponds to a time series sample of a state monitoring channel;

[0094] S32. Taking the data matrix to be encoded as input, combining it with the dynamic sparse representation dictionary corresponding to the sliding window, using the orthogonal matching pursuit algorithm to perform sparse encoding on each state observation data, solving the sparse coefficient vector α, satisfying the following expression: x≈D·α, where x is the state observation vector, D is the dynamic sparse representation dictionary, and α is the sparse coding coefficient vector, and extracting the state sparse feature vector;

[0095] S33. Input the sparse coding coefficient vector α into the state prediction structure. The state prediction structure is composed of two nonlinear mapping functions, including a gate activation function and a state projection function:

[0096]

[0097] in, is the current state prediction value, W g 、W p are the weight matrices of the gated path and the predicted path, b g 、b p is the corresponding bias term, σ(·) is the nonlinear activation function;

[0098] The formula demonstrates that the state prediction value is composed of the product of the results of two independent paths: one path modulates the response to sparse features through a nonlinear activation function, and the other path obtains the basic prediction through linear projection. The gating path is responsible for adjusting the activation level of different features in the current state, while the projection path provides the original state mapping output. The fusion of the two effectively captures the nonlinear relationship between features, enables precise prediction of complex dynamic states, and improves the ability to model subtle state evolution trends.

[0099] S34. Arrange the state prediction values ​​generated by the continuous sliding windows in chronological order to construct a state prediction sequence.

[0100] This embodiment extracts multi-channel state observation data within a sliding window to construct a data matrix to be encoded, and combines it with a dynamic sparse representation dictionary to perform sparse encoding on the state observation vector using an orthogonal matching pursuit algorithm to obtain a state sparse feature vector. This feature vector is input into a nonlinear state prediction structure composed of a gated activation function and a state projection function to generate a current state prediction value and construct a state prediction sequence. This method introduces sparse reconstruction and nonlinear prediction mechanisms in the state modeling process, which not only retains the key dynamic change characteristics of the rocket's erection state, but also avoids the risks of overfitting and misestimation, significantly improving the state prediction accuracy under complex disturbance conditions and the ability to respond to small anomalies.

[0101] In this embodiment, the S4 specifically includes:

[0102] S41. Obtain the current state prediction value in the constructed state prediction sequence And the actual state observation value y at the corresponding moment t , as the current predicted value and the current observed value respectively;

[0103] S42. Predicted value based on current state and the actual state observation value y t The difference between the residual signal ε is calculated t ;

[0104] S43, based on the residual signal sequence {ε t-n ,…,ε t}, calculate the residual change rate Δε t , the residual change rate is the difference between the residual signal at the current moment and the residual signal at the previous moment;

[0105] S44, according to the residual signal ε t and the residual change rate Δε t , construct the feedback regulation function f(·), and calculate the feedback regulation coefficient γ t , the feedback adjustment function adopts the form of exponential weight combination, which is expressed as follows:

[0106] γ t =exp(-λ1|ε t |-λ2|Δε t |);

[0107] Among them, γ t is the feedback adjustment coefficient, λ1 and λ2 are the adjustment weight coefficients of the residual amplitude and change rate, respectively, which are used to control the response degree to the residual intensity and change trend.

[0108] The formula uses an exponential decay structure, dynamically generating an adjustment coefficient between zero and one based on the amplitude of the current residual signal and the rate of change of the residual. This coefficient is used to control the amplitude and direction of subsequent parameter adjustments, achieving synchronous response adjustment of the state prediction path and the disturbance compensation path. The exponential structure in the formula effectively suppresses the dramatic fluctuations of the adjustment coefficient caused by outliers, making the adjustment behavior smooth and gradual, with strong numerical stability and dynamic adaptability. This ensures that parameter updates are sufficiently sensitive in the early stages of an anomaly and gradually converge as the system stabilizes, avoiding overcorrection.

[0109] This embodiment constructs a residual signal by obtaining the difference between the current state prediction value and the actual state observation value, and further calculates the residual change rate based on the change of the residual signal in a continuous time period to form a state deviation feature from the perspective of time evolution. On this basis, the residual signal and the residual change rate are used as input to construct an exponential weighted feedback adjustment function to generate a feedback adjustment coefficient. This feedback adjustment coefficient serves as the core variable of the system's adaptive control and jointly affects the adjustment of subsequent filter gain parameters and disturbance modeling parameters. By introducing a dynamic combination mechanism of residual information intensity and trend changes, a sensitive response and real-time feedback to the abnormal offset evolution process are achieved, which enhances the system's recognition accuracy and control flexibility for nonlinear offsets and disturbance mutations during the erection of offshore rockets.

[0110] In this embodiment, the S5 specifically includes:

[0111] S51. Receive the state sparse feature vector, use the state sparse feature vector as the input of the state prediction path, and use the prediction model in the Kalman filter structure to perform state estimation to generate a state estimation value:

[0112]

[0113] in, is the estimated value of the current state, is the state estimate at the previous moment, A is the state transfer matrix, B is the control input matrix, u t is an optional control input;

[0114] The formula reflects the evolution of the system state over time. The state transition matrix describes the state transition characteristics of the system under undisturbed conditions, and the control input matrix reflects the impact of external inputs on state changes. This prediction process provides an initial estimate for subsequent state updates and error correction, ensuring the temporal consistency and recursiveness of the entire filtering process. It is the foundation for state tracking and continuous estimation.

[0115] S52: Receive the residual signal, use the residual signal as the input of the disturbance compensation path, and use the exponentially weighted residual modeling method to perform disturbance estimation to generate a disturbance estimation value:

[0116]

[0117] Among them, d t is the current disturbance estimate, ε t-i is the residual signal before the i-th step, ω i Estimate the weight coefficient for the disturbance to meet the normalization condition;

[0118] This method assigns different weights to the residual signals at multiple historical time steps, forming a weighted sum of the disturbance estimates. The weight coefficients satisfy normalization constraints. The formula exploits the principle of temporal proximity, making the residual signals of more recent time steps have a greater influence on the current disturbance estimate. This enhances the model's responsiveness to sudden disturbances and improves the accuracy of the disturbance trend while maintaining estimated stability.

[0119] S53. Perform weighted fusion on the state estimation value generated by the state prediction path and the disturbance estimation value generated by the disturbance compensation path to construct a final state estimation value. The fusion method adopts a weighted average model, which is expressed as:

[0120]

[0121] in, is the current final state estimate, and λ is the fusion weight coefficient, which balances the state estimation and disturbance estimation.

[0122] The formula implements the numerical fusion processing between the predicted state value output by the state prediction path and the disturbance estimate output by the disturbance compensation path. This fusion method allocates the weight ratio between the two path outputs by setting a fusion weight coefficient, and generates the final state estimate value according to the linear combination method. The principle of the formula is that when the system state is less affected by the disturbance, the output of the state prediction path accounts for a relatively higher proportion; when the disturbance signal is more significant, the estimated result of the disturbance compensation path occupies a dominant position in the fusion, thus achieving a balanced mechanism for simultaneously modeling and responding to state changes and disturbance characteristics. Through a simple and effective linear fusion strategy, this method realizes the integration of multi-path information without the need for complex calculations, thereby improving the stability and expression integrity of the state estimation results.

[0123] This implementation inputs the state sparse feature vector into the state prediction path, predicting the current erection state based on a Kalman filter structure. Simultaneously, the residual signal is input into the disturbance compensation path, estimating the impact of disturbances using an exponentially weighted model. The two paths generate state and disturbance estimates, respectively, which are then weightedly fused to produce the final state estimate. This method achieves decoupled modeling of state evolution and external disturbances, and effectively balances their impact on the estimation results through a fusion strategy. This improves state tracking accuracy and system stability in complex sea conditions, providing a reliable foundation for subsequent anomaly detection.

[0124] In this embodiment, S6 specifically includes:

[0125] S61, receiving a feedback adjustment coefficient, and inputting the feedback adjustment coefficient into the state prediction path and the disturbance compensation path respectively, for linkage adjustment of the filter gain parameter in the state estimation path and the disturbance modeling parameter in the disturbance compensation path;

[0126] S62. In the state prediction path, dynamically update a filter gain parameter based on the current value of the feedback adjustment coefficient. The filter gain parameter is used in the state update step in the Kalman filter structure. When the feedback adjustment coefficient increases, the filter gain parameter is correspondingly increased to enhance the ability to track rapid state changes. When the feedback adjustment coefficient decreases, the filter gain parameter is correspondingly decreased to suppress invalid responses.

[0127] S63. In the disturbance compensation path, adjusting disturbance modeling parameters according to a change trend of the current feedback adjustment coefficient, wherein the disturbance modeling parameters include a time-weighted window length for disturbance estimation and a weighted distribution structure of historical residual signals. When the feedback adjustment coefficient is increased, the proportion of the current residual signal in the disturbance estimation is increased, and when the feedback adjustment coefficient is decreased, the influence of historical disturbance trend information is increased.

[0128] S64. Within the current sliding window period, respectively calculate the state estimation error between the state estimation value output by the state prediction path and the actual state observation value, and the disturbance estimation error between the disturbance estimation value output by the disturbance compensation path and the actual state observation value, construct a joint error evaluation index, and establish a parameter adjustment criterion with the goal of minimizing the joint error;

[0129] S65. Based on the output result of the joint error evaluation index, dynamically adjust the adjustment step size and update direction of the filter gain parameter used in the state prediction path and the disturbance modeling parameter used in the disturbance compensation path, and establish a linked coupled dual-factor parameter adaptive adjustment mechanism;

[0130] S66. Apply the filter gain parameters and disturbance modeling parameters updated by the joint error drive to the execution process of the state prediction path and the disturbance compensation path at the next moment, forming a continuous parameter self-optimization closed-loop system driven by the feedback adjustment coefficient.

[0131] This implementation method adjusts the filter gain parameters in the state prediction path and the disturbance modeling parameters in the disturbance compensation path based on the feedback adjustment coefficient. By introducing a joint error evaluation index, a dynamic trade-off mechanism between the state estimation error and the disturbance estimation error is established, thus realizing the linkage adjustment and coupling control of the dual-path parameters. The parameter adjustment step size and update direction are corrected in real time within the sliding window cycle, and the updated parameters are continuously applied to the state estimation and disturbance compensation process of the next cycle, thus constructing a feedback-driven parameter self-optimization closed-loop system. This mechanism can dynamically adapt to system changes under conditions of multi-source disturbances and non-stationary states, improve the stability of state estimation and the ability to respond to abnormal trends, and effectively enhance the adaptability and robustness of the state perception system during the rocket erection process.

[0132] In this embodiment, S65 specifically includes:

[0133] S651. Generate a joint error evaluation index based on the calculated state estimation error and disturbance estimation error, where the joint error evaluation index is composed of a weighted combination of the state estimation error and the disturbance estimation error.

[0134] S652: Compare the joint error evaluation index with a preset joint error threshold, and determine whether to initiate a parameter adjustment process based on the comparison result. If the joint error evaluation index is greater than the preset joint error threshold, perform parameter adjustment; if not, maintain the current parameter settings.

[0135] S653: In parameter adjustment, the change value of the state estimation error is called as an adjustment control variable and input into the state prediction path to adjust the filter gain parameter. The adjustment operation sets the adjustment step size and direction according to the error change value to update the filter gain parameter;

[0136] S654: Synchronously call the change value of the disturbance estimation error and transmit it to the disturbance compensation path as a control input to adjust the disturbance modeling parameters. The disturbance modeling parameter adjustment strategy includes adjusting the time-weighted distribution of the residual signal and the estimation window span;

[0137] S655: Jointly analyze the error change value of the state prediction path and the error change value of the disturbance compensation path to calculate a joint adjustment weighting factor. The joint adjustment weighting factor is used to allocate adjustment resources between the two paths.

[0138] S656: Output the updated filter gain parameters and disturbance modeling parameters respectively and pass them to S66 to build a continuous dual-path parameter linkage update chain.

[0139] This embodiment extracts the state estimation error and the disturbance estimation error, constructs a joint error evaluation index, and combines the preset threshold to determine whether to enter the parameter adjustment stage, thereby realizing adaptive control based on error dynamic feedback. During the adjustment process, the change values ​​of the state estimation error and the disturbance estimation error are respectively called to control the update step size and direction of the filter gain parameter and the disturbance modeling parameter. At the same time, a joint adjustment weight factor is introduced to configure parameter adjustment resources between the state prediction path and the disturbance compensation path. This method forms a continuous dual-path parameter linkage update chain, which not only realizes the coordinated adjustment of state estimation and disturbance compensation, but also improves the system's ability to adapt to parameter changes in complex disturbance environments, and enhances the response accuracy and stability to abnormal conditions.

[0140] Example 1:

[0141] In order to verify the feasibility of the present invention in implementation, the present invention was applied to three sea-based rocket vertical erection missions carried out by a coastal commercial launch platform in China between August 2024 and February 2025. The test environment was located in the Yellow Sea. Since the launch platform is affected by wind and wave interference and structural micro-vibration response for a long time, there are significant uncertainty and delay problems in state monitoring during the rocket erection process. Conventional anomaly detection methods based on static thresholds and fixed filtering strategies cannot meet the real-time and accuracy requirements. To this end, the system of the present invention is deployed in the platform state perception and control module, and real-time processing is performed in combination with 12 sets of multi-source inertial navigation, inclination, strain and acceleration sensor data installed on the platform.

[0142] During actual deployment, the system first segments each set of sensor data into time series using a dynamic sliding window mechanism, automatically adjusting the window length and generating a multi-channel sparse feature dictionary. Feature extraction is then performed using an orthogonal matching pursuit algorithm. The current rocket attitude is then estimated using a gated nonlinear prediction structure, and feedback adjustment coefficients are established in conjunction with real-time residual signals. In the dual-path filtering structure, the state prediction path uses state sparse features for Kalman prediction, while the disturbance compensation path constructs disturbance estimates based on the residuals and fuses them. The system then uses error feedback to coordinately adjust the filter gain and disturbance modeling parameters, ultimately outputting the estimated state and driving the anomaly detection module.

[0143] In order to fully evaluate the performance of the system, we selected a vertical erection test of a medium-sized liquid carrier rocket in October 2024 as a detailed test sample. During the test, the total erection time of the rocket was 540 seconds, and the erection angle ranged from horizontal 0° to vertical 90°. Between 120 and 160 seconds after erection, the platform encountered wind level 5 side wave interference, resulting in significant disturbances in the structural response. The traditional state estimation system failed to identify the abnormal trend in time and only discovered the deviation after the erection was completed. The system of the present invention detected the attitude estimation residual abnormality at the 128th second, triggered the status warning in real time, and output the judgment signal about 370 seconds in advance. The following table shows the performance comparison of the method of the present invention and the traditional method in three sea rocket erection missions:

[0144] Table 1 Comparison of condition monitoring performance based on the method of the present invention and the traditional method

[0145]

[0146]

[0147] The data shows that the proposed system improves anomaly detection response time by nearly 6.6 times compared to traditional methods, effectively avoiding the risk of delayed response. In terms of state estimation accuracy, the maximum angle error is reduced from 3.8° to 1.1°, and the root mean square error is reduced by approximately 65%. Furthermore, the false alarm rate and missed detection rate are significantly reduced, improving the overall reliability and stability of the system.

[0148] In a specific scenario, the system successfully handled a low-temperature, high-humidity sea test in November 2024. During this mission, some strain data was lost due to temporary moisture in the sensor interface. The system, using a collaborative compressed sensing mechanism, reconstructed the lost data and maintained a continuous and stable state estimate, preventing the rocket control system from misjudging the erection state and ensuring the safe completion of the mission.

[0149] In addition, during the operation of the system, the residual-driven linkage adjustment mechanism designed in the present invention adaptively adjusts the filtering and modeling parameters in multiple disturbance periods. Without human intervention, it can autonomously adjust the adjustment intensity of the state prediction path and the disturbance compensation path according to the joint error index, so that the filtering error quickly converges to below 0.5° within 120 seconds and maintains stable estimation capabilities in subsequent stages, fully demonstrating the effectiveness of the closed-loop feedback and dynamic adjustment structure.

[0150] In summary, this embodiment demonstrates the robustness and accuracy of the method of the present invention in complex disturbance environments, especially in the offshore platform erection status perception scenario. It can respond to sudden disturbances, information loss and status anomalies in real time, and build a complete data perception-state prediction-disturbance compensation-feedback adjustment-abnormality identification closed-loop process, providing a stable, efficient and deployable status monitoring technology solution for offshore aerospace erection missions.

[0151] The above description is only a preferred specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any technician familiar with the technical field, within the technical scope disclosed by the present invention, who makes equivalent replacements or changes based on the technical solution and inventive concept of the present invention, should be covered by the scope of protection of the present invention.

Claims

1. A method for detecting abnormalities in the erection state of a sea-based rocket based on sparse representation and adaptive filtering, characterized in that: The steps include: S1. Obtain the original state observation data collected by multiple state monitoring sensors during the erection process of the sea rocket, perform time alignment processing, and generate a multi-channel state observation time series sequence; S2. Construct a dynamic sliding window mechanism based on the multi-channel state observation time series, set an adaptive window length for each dynamic sliding window, and generate a dynamic sparse representation dictionary; S3. Sparsely encode the state observation data in the dynamic sliding window based on the dynamic sparse representation dictionary, extract the state sparse feature vector, and calculate the current state prediction value based on the state sparse feature vector; S4. Calculate the residual signal between the current state prediction value and the actual state observation value, and calculate the feedback adjustment coefficient based on the residual signal; S5. Construct a dual-path filtering architecture including a state prediction path and a disturbance compensation path, wherein the state prediction path receives a state sparse feature vector for performing state estimation, and the disturbance compensation path receives a residual signal for performing disturbance estimation, and the outputs of the dual paths are fused to generate a final state estimation result; S6. According to the feedback adjustment coefficient, the filter gain parameter used in state estimation and the disturbance modeling parameter used in disturbance estimation are adjusted in a coordinated manner to establish a parameter adaptive adjustment mechanism; S7. Input the final state estimation result and the residual signal into an anomaly detection model. The anomaly detection model identifies state anomalies during the erection of the sea-based rocket based on a time series consistency analysis method and outputs an anomaly labeling result.

2. The method for detecting abnormalities in the erection state of a sea-based rocket based on sparse representation and adaptive filtering according to claim 1, wherein: The S2 specifically includes: S21, performing unified time axis mapping on the multi-channel state observation time series to construct a multi-channel observation matrix covering the synchronous data of each channel; S22. In the multi-channel observation matrix, set the initial position of the sliding window, and extract the state observation data of each channel in the corresponding time period as the data input segment corresponding to the current sliding window; S23. For the data input segment corresponding to the current sliding window, calculate the state change rate of each channel in the time period, and construct a nonlinear window length adjustment function based on the average state change rate of all channels. Define the current window length L(t) as: Among them, L min is the minimum length of the sliding window, L max is the maximum length of the sliding window, v i (t) is the state change rate of the i-th channel in the current time period, and N is the total number of state monitoring channels; S24. Calculate the sliding residual signal between the state prediction value and the actual state observation value in the previous sliding cycle of the current sliding window, and calculate the change amplitude Δε of the sliding residual signal. If the change amplitude Δε is greater than the set residual change threshold θ, move the starting position of the sliding window forward by a sliding step Δt, and update the sliding window position. If Δε is less than or equal to the residual change threshold θ, keep the starting position of the sliding window unchanged. S25. Extract observation data segments within each triggered sliding window, and construct a set of dynamic sparse representation dictionaries corresponding to the time features of each sliding window.

3. The method for detecting abnormalities in the erection state of a sea-based rocket based on sparse representation and adaptive filtering according to claim 1, wherein: The S3 specifically includes: S31, extracting state observation data segments within the time period corresponding to each sliding window, and constructing a data matrix to be encoded; S32. Taking the data matrix to be encoded as input, combining it with the dynamic sparse representation dictionary corresponding to the sliding window, using the orthogonal matching pursuit algorithm to perform sparse encoding on each state observation data, solving the sparse coefficient vector α, satisfying the following expression: x≈D·α, where x is the state observation vector, D is the dynamic sparse representation dictionary, and α is the sparse coding coefficient vector, and extracting the state sparse feature vector; S33. Input the sparse coding coefficient vector α into the state prediction structure. The state prediction structure is composed of two nonlinear mapping functions, including a gate activation function and a state projection function: in, is the current state prediction value, W g 、W p are the weight matrices of the gated path and the predicted path, b g 、b p is the corresponding bias term, σ(·) is the nonlinear activation function; S34. Arrange the state prediction values ​​generated by the continuous sliding windows in chronological order to construct a state prediction sequence.

4. The method for detecting abnormalities in the erection state of a sea-based rocket based on sparse representation and adaptive filtering according to claim 1, wherein: The S4 specifically includes: S41. Obtain the current state prediction value in the constructed state prediction sequence And the actual state observation value y at the corresponding moment t , as the current predicted value and the current observed value respectively; S42. Predicted value based on current state and the actual state observation value y t The difference between the residual signal ε is calculated t ; S43, based on the residual signal sequence {ε t-n ,…,ε t }, calculate the residual change rate Δε t , the residual change rate is the difference between the residual signal at the current moment and the residual signal at the previous moment; S44, according to the residual signal ε t and the residual change rate Δε t , construct the feedback regulation function f(·), and calculate the feedback regulation coefficient γ t , the feedback adjustment function adopts the form of exponential weight combination.

5. The method for detecting abnormalities in the erection state of a sea-based rocket based on sparse representation and adaptive filtering according to claim 1, wherein: The S5 specifically includes: S51, receiving the state sparse feature vector, using the state sparse feature vector as the input of the state prediction path, and using the prediction model in the Kalman filter structure to perform state estimation to generate a state estimation value S52, receiving the residual signal, taking the residual signal as the input of the disturbance compensation path, and using the exponentially weighted residual modeling method to perform disturbance estimation to generate a disturbance estimation value d t ; S53. Perform weighted fusion on the state estimation value generated by the state prediction path and the disturbance estimation value generated by the disturbance compensation path to construct a final state estimation value. The fusion method adopts a weighted average model.

6. The method for detecting abnormalities in the erection state of a sea-based rocket based on sparse representation and adaptive filtering according to claim 1, wherein: The S6 specifically includes: S61, receiving a feedback adjustment coefficient, and inputting the feedback adjustment coefficient into the state prediction path and the disturbance compensation path respectively, for linkage adjustment of the filter gain parameter in the state estimation path and the disturbance modeling parameter in the disturbance compensation path; S62. In the state prediction path, dynamically update the filter gain parameter according to the value of the current feedback adjustment coefficient. The filter gain parameter is used in the state update step in the Kalman filter structure. S63. In the disturbance compensation path, adjusting disturbance modeling parameters according to a change trend of the current feedback adjustment coefficient, wherein the disturbance modeling parameters include a time-weighted window length of the disturbance estimation and a weighted distribution structure of the historical residual signal; S64. Within the current sliding window period, respectively calculate the state estimation error between the state estimation value output by the state prediction path and the actual state observation value, and the disturbance estimation error between the disturbance estimation value output by the disturbance compensation path and the actual state observation value, and construct a joint error evaluation index; S65. Based on the output result of the joint error evaluation index, dynamically adjust the adjustment step size and update direction of the filter gain parameter used in the state prediction path and the disturbance modeling parameter used in the disturbance compensation path, and establish a linked coupled dual-factor parameter adaptive adjustment mechanism; S66. Apply the filter gain parameters and disturbance modeling parameters updated by the joint error drive to the execution process of the state prediction path and the disturbance compensation path at the next moment, forming a continuous parameter self-optimization closed-loop system driven by the feedback adjustment coefficient.

7. The method for detecting abnormalities in the erection state of a sea-based rocket based on sparse representation and adaptive filtering according to claim 6, characterized in that: S65 specifically includes: S651. Generate a joint error evaluation index based on the calculated state estimation error and disturbance estimation error, where the joint error evaluation index is composed of a weighted combination of the state estimation error and the disturbance estimation error. S652: Compare the joint error evaluation index with a preset joint error threshold, and determine whether to initiate a parameter adjustment process based on the comparison result. If the joint error evaluation index is greater than the preset joint error threshold, perform parameter adjustment; if not, maintain the current parameter settings. S653: In parameter adjustment, the change value of the state estimation error is called as an adjustment control variable and input into the state prediction path to adjust the filter gain parameter. The adjustment operation sets the adjustment step size and direction according to the error change value to update the filter gain parameter; S654: Synchronously call the change value of the disturbance estimation error and transmit it to the disturbance compensation path as a control input to adjust the disturbance modeling parameters. The disturbance modeling parameter adjustment strategy includes adjusting the time-weighted distribution of the residual signal and the estimation window span; S655: Jointly analyze the error change value of the state prediction path and the error change value of the disturbance compensation path to calculate a joint adjustment weighting factor. The joint adjustment weighting factor is used to allocate adjustment resources between the two paths. S656: Output the updated filter gain parameters and disturbance modeling parameters respectively and pass them to S66 to build a continuous dual-path parameter linkage update chain.

Images