Multidimensional Time Series Data Diagnosis and Completion Method, Device, System and Medium

Abstract

The present invention relates to the technical field of industrial production abnormal data diagnosis, and discloses a multi-dimensional time series data diagnosis and completion method, device, system and medium. The method includes collecting sensor time series data, constructing a neural network model, defining a loss function of the neural network model and training it, inputting the multi-dimensional time series data obtained in actual production into the trained neural network model to obtain reconstructed multi-dimensional time series data; calculating the observation error and updating the mask matrix, and retraining the neural network model; until no new outliers can be found or a specific number of cycles is reached, missing value completion is achieved. During the reconstruction process, the original time series is recursively synthesized by inverse wavelet transform, and the reconstruction errors of subtle anomalies will be accumulated and amplified, thereby helping to discover the abnormal parts of the data; during the reconstruction, it can simultaneously complete its own missing parts and the missing parts removed due to data anomalies.

Description

Technical Field

[0001] The present invention relates to the technical field of industrial production abnormal data diagnosis, and specifically relates to a method, device, system and medium for diagnosing and completing multi-dimensional time series data. Background Technique

[0002] In industrial production, by deploying different types of sensors in the system, a continuously generated data stream can be obtained, generally referred to as time series data. In practice, there are problems such as sensor failures or damages, data transmission interruptions or losses, human operation errors, equipment maintenance, shutdown repairs, and environmental mutations, resulting in phenomena such as random continuous missing and sequence change anomalies in multi-source time series data. If not solved, it may lead to a decline in the robustness of the later analysis model training, making it more sensitive to fluctuations and anomalies in the input data, thus affecting the performance and generalization ability of the model.

[0003] In terms of anomaly diagnosis, existing technologies mostly build anomaly recognition models based on manually given anomaly labels. However, in practice, it is difficult to obtain anomaly labels, and the process of obtaining anomaly labels consumes a large amount of manpower and material resources. In terms of data completion, interpolation methods are mostly used, without considering the long-term periodicity of the data and the correlation between different variables (sensors). Summary of the Invention

[0004] To solve the above technical problems, the present invention proposes a method, device, system and medium for diagnosing and completing multi-dimensional time series data, integrating wavelet transform into the network and reconstructing different frequency bands respectively; introducing autocorrelation attention to evaluate the periodicity within the sequence; introducing asymmetric graph convolution to capture the correlation between variables; reconstructing and taking the difference to identify anomalies, and using masked supervision to complete the missing data.

[0005] To solve the above technical problems, the present invention adopts the following technical solutions:

[0006] A method for diagnosing and completing multi-dimensional time series data specifically includes the following steps:

[0007] Step 1, collect the sensor time series data generated by n sensors on the production equipment to obtain multi-dimensional time series data where x i represents the sensor time series data generated by the i-th sensor, 1 ≤ i ≤ n; is the w-th sensor data in x i and w is the length of the sensor time series data of each sensor;

[0008] Step 2, construct a neural network model: perform the first-level multi-wavelet basis function discrete wavelet transform on each x i in X respectively to obtain x iThe corresponding first-level transformed low-frequency components and the first-level transformed high-frequency components are respectively input into a one-dimensional convolutional network and a fully-connected network to obtain the first-level fused low-frequency sequence and the first-level fused high-frequency sequence Perform a second-level multi-wavelet basis function discrete wavelet transform to obtain the second-level transformed low-frequency components and the second-level transformed high-frequency components are respectively input into a one-dimensional convolutional network and a fully-connected network to obtain the second-level fused low-frequency sequence and the second-level fused high-frequency sequence are and input into a self-correlation fusion module composed of self-correlation attention, residual connection, and a one-dimensional convolutional network to obtain the first-level time series feature map of the i-th sensor are and input into the self-correlation fusion module to obtain the second-level time series feature map of the i-th sensor

[0009] are input into a one-dimensional convolutional network after the aggregation result obtained by inputting into a graph convolutional network to obtain a first-level high-frequency sequence are input into a one-dimensional convolutional network after the aggregation result obtained by inputting into a graph convolutional network to obtain a second-level low-frequency sequence and a second-level high-frequency sequence Perform and a multi-wavelet basis function wavelet inverse transform to obtain a first-level low-frequency sequence By and performing a multi-wavelet basis function wavelet inverse transform to obtain the reconstructed data By taking the column-wise average to obtain the reconstructed time series data of the i-th sensor and further obtain the reconstructed multi-dimensional time series data

[0010] Step 3, define the loss function of the neural network model and perform training; the loss function is

[0011]

[0012] where, M Xis the mask matrix corresponding to X; if X is filled with zero due to missing at a specific position, the element of M at the specific position is 0, and in other cases, the element of M at the specific position is 1; M H , M LH , M LL are respectively the mask matrices corresponding to ; ° represents element-wise multiplication; λ X , λ H , λ LH , λ LL represents the weighting coefficient, ‖·‖2 represents the L2 norm;

[0013] Step Four, anomaly diagnosis and code completion, specifically including the following steps:

[0014] Step S41, input the multi-dimensional time series data X0 obtained in actual production into the trained neural network model to obtain the reconstructed multi-dimensional time series data Calculate the observation error Δ: If the observation error Δ exceeds the set threshold, the corresponding sensor data in X0 is removed as an outlier and filled with zero, and the mask matrix is updated, and the neural network model is retrained; the updated multi-dimensional time series data X0 is input into the retrained neural network model;

[0015] Step S42, loop and run Step S41 until no new outliers can be found or a specific number of loops is reached. Finally, the has removed all abnormal sensor data in the multi-dimensional time series data X0 and completed the missing value filling.

[0016] Furthermore, the wavelet basis functions used in the multi-wavelet basis function discrete wavelet transform include two or more of Haar wavelet, Daubechies wavelet, Mexican Hat wavelet, Morlet wavelet, and Meyer wavelet.

[0017] Furthermore, in Step Two, for each x in X i perform the first-level multi-wavelet basis function discrete wavelet transform to obtain the first-level transform low-frequency component i and the first-level transform high-frequency component of x Input into the one-dimensional convolutional network and the fully connected network respectively to obtain the first-level fusion low-frequency sequence and the first-level fusion high-frequency sequence Specifically including:

[0018] The time series data x of the i-th sensor iAfter the discrete wavelet transform with multi-wavelet basis functions, the low-frequency component of the first-level transform is obtained and the high-frequency component of the first-level transform

[0019]

[0020] where mbDWT(·) represents the discrete wavelet transform with multi-wavelet basis functions, and k is the total number of wavelet basis functions in the discrete wavelet transform with multi-wavelet basis functions; represents the discrete wavelet transform low-frequency component of x i under the j-th wavelet basis function, 1 ≤ j ≤ k;

[0021] For perform fusion to obtain the first-level fusion low-frequency sequence Specifically, use a multi-core one-dimensional convolutional network 1DCNN to transform the discrete wavelet transform low-frequency component corresponding to each wavelet basis function perform transformation, input the transformation result into a fully connected network FC for fusion, and obtain the first-level fusion low-frequency sequence

[0022]

[0023] Input the high-frequency component of the first-level transform into a multi-core one-dimensional convolutional network 1DCNN and a fully connected network FC to obtain the first-level fusion high-frequency sequence

[0024]

[0025] Furthermore, in step two, the process of inputting and into a self-correlation fusion module composed of self-correlation attention, residual connection, and one-dimensional convolutional network to obtain the first-level time series feature map of the i-th sensor Input and into the self-correlation fusion module to obtain the second-level time series feature map of the i-th sensor Specifically include:

[0026] Input the first-level fusion low-frequency sequence of the i-th sensor and the first-level fusion high-frequency sequence into the self-correlation attention AC respectively and perform residual connection, input the obtained result into a one-dimensional convolutional network 1DCNN and then merge to obtain the first-level time series feature map of the i-th sensor

[0027]

[0028] Among them, d is the number of kernels in the one-dimensional convolutional network, and || represents merging;

[0029] Input the second-level fusion low-frequency sequence and the second-level fusion high-frequency sequence into the autocorrelation attention AC respectively and perform residual connection. Input the obtained results into the one-dimensional convolutional network 1DCNN and then merge them to obtain the second-level time series feature map of the i-th sensor

[0030]

[0031] A multi-dimensional time series data diagnosis and completion device, comprising:

[0032] A data collection module that collects sensor time series data generated by n sensors on production equipment to obtain multi-dimensional time series data Among them, x i represents the sensor time series data generated by the i-th sensor, where 1 ≤ i ≤ n; is the w-th sensor data in x i , and w is the length of the sensor time series data of each sensor;

[0033] A neural network model construction module that performs the first-level multi-wavelet basis function discrete wavelet transform on each x in X respectively to obtain the first-level transform low-frequency component i corresponding to x i and the first-level transform high-frequency component and the first-level transform high-frequency component Input into the one-dimensional convolutional network and the fully connected network respectively to obtain the first-level fusion low-frequency sequence and the first-level fusion high-frequency sequence Perform the second-level multi-wavelet basis function discrete wavelet transform on to obtain the second-level transform low-frequency component and the second-level transform high-frequency component Input into the one-dimensional convolutional network and the fully connected network respectively to obtain the second-level fusion low-frequency sequence and the second-level fusion high-frequency sequence Input and into the autocorrelation fusion module composed of autocorrelation attention, residual connection and one-dimensional convolutional network to obtain the first-level time series feature map of the i-th sensor Input and Input it into the autocorrelation fusion module to obtain the second-level time series feature map of the i-th sensor

[0034] Input the aggregation result obtained after inputting it into the graph convolutional network into a one-dimensional convolutional network to obtain a first-level high-frequency sequence Input the aggregation result obtained after inputting it into the graph convolutional network into a one-dimensional convolutional network to obtain a second-level low-frequency sequence and a second-level high-frequency sequence Perform inverse multi-wavelet basis function wavelet transformation on and to obtain a first-level low-frequency sequence By performing inverse multi-wavelet basis function wavelet transformation on and to obtain the reconstructed data By taking the average of column by column to obtain the reconstructed time series data of the i-th sensor Furthermore, obtain the reconstructed multi-dimensional time series data

[0035] Training module, define the loss function of the neural network model and perform training; the loss function is: where M X is the mask matrix corresponding to X; if X is filled with zeros due to missing values at specific positions, the elements of M at the specific positions are 0; M H , M LH , M LL are the mask matrices corresponding to respectively; ° represents element-wise multiplication; λ X , λ H , λ LH , λ LL represent the weighting coefficients, and ‖·‖2 represents the second norm;

[0036] Anomaly diagnosis and code completion module, input the multi-dimensional time series data X0 obtained in actual production into the trained neural network model to obtain the reconstructed multi-dimensional time series data Calculate the observation error Δ: If the observation error Δ exceeds the set threshold, then the corresponding sensor data in X0 is removed as an outlier and filled with zeros, and the mask matrix is updated, and the neural network model is retrained; input the updated multi-dimensional time series data X0 into the retrained neural network model; loop the above content until no new outliers can be found or a specific number of loops is reached, and finally obtain the All abnormal sensor data in the multi-dimensional time series data X0 have been removed and missing value filling has been implemented.

[0037] A multi-dimensional time series data diagnosis and completion system includes a memory and a processor. The memory stores a computer program. When the computer program is executed by the processor, the processor executes the steps of the method.

[0038] A computer-readable storage medium stores a computer program. When the computer program is executed by a processor, the processor executes the steps of the method.

[0039] Compared with the prior art, the beneficial technical effects of the present invention are:

[0040] In the process of reconstruction of the present invention, the inverse wavelet transform recursively synthesizes the original time series, and the reconstruction errors of minor anomalies will be accumulated and amplified, thereby helping to discover the abnormal parts of the data; during reconstruction, it can simultaneously fill in its own missing parts and the missing parts removed due to data anomalies. Description of the Drawings

[0041] Figure 1 It is a schematic diagram of the neural network model adopted by the present invention. Detailed Embodiment

[0042] The following will make a detailed description of a preferred embodiment of the present invention with reference to the drawings.

[0043] This embodiment provides a method for diagnosing and completing multi-dimensional time series data by unsupervised deep learning, which is specifically introduced as follows.

[0044] 1. Organize industrial site data:

[0045] Since there are no manual labels, self-supervised learning needs to be used to reconstruct the time series by modeling the time dependence and the dependence between variables for anomaly detection, repair, and completion. Let the n-dimensional original multi-dimensional time series data be x i representing the sensor time series data generated by the i-th sensor, and its missing parts are filled with 0, and w is the length or window width of each sensor time series data.

[0046] 2. Design a neural network model:

[0047] The discrete wavelet transform is separately performed on the sensor time series data of each dimension in the multi-dimensional time series data X to obtain the corresponding low-frequency component and high-frequency component. It should be noted that since it is uncertain which wavelet basis function is optimal, the present invention adopts the multi-basis discrete wavelet transform (Multi-basis DWT, mbDWT), such as Haar wavelet, Daubechies wavelet, etc. Specifically, the i-th sensor time series data x i After the multi-basis discrete wavelet transform, the first-level transform low-frequency component is obtained And the first-level transform high-frequency component That is:

[0048]

[0049] Among them, Each row of represents the discrete wavelet transform low-frequency component of x i Under a certain wavelet basis function, it characterizes the macroscopic trend of the time series; Each row of represents the discrete wavelet transform high-frequency component of x i Under a certain wavelet basis function, it characterizes the microscopic details of the time series, and k is the total number of wavelet basis functions. It should be noted that: the discrete wavelet transform includes filtering and downsampling. Each time the discrete wavelet transform is performed, the length of the sequence becomes half of the original sequence. Before further transforming Considering that the multi-basis discrete wavelet transform of the k-dimensional sequence will obtain a k 2 Dimensional sequence. If p-level decomposition is performed, a Dimensional sequence, significantly increasing unnecessary model complexity and computational workload. Therefore, it is necessary to perform fusion on To obtain the first-level fusion low-frequency sequence Specifically, a multi-core one-dimensional convolutional network 1DCNN can be used to perform transformation on the discrete wavelet transform low-frequency component corresponding to each wavelet basis function To obtain the corresponding multi-dimensional feature sequence, and then use the fully connected network FC for fusion to obtain the one-dimensional first-level fusion low-frequency sequence That is:

[0050]

[0051] For the high-frequency component Although there is no need to continue decomposition, the same method can also be used for transformation to reduce the number of parameters, and the first-level fusion high-frequency sequence That is:

[0052]

[0053] The second-level transform is similar to the first-level. The Perform the discrete wavelet transform with multi-wavelet basis functions to obtain the low-frequency component of the second-level transform and the high-frequency component of the second-level transform That is:

[0054]

[0055] And perform similar fusion to obtain the low-frequency sequence of the second-level fusion and the high-frequency sequence of the second-level fusion In the above symbols, the subscripts L and H represent low frequency and high frequency respectively. The second-level decomposition is to decompose the low-frequency component of the first-level decomposition again. Therefore, the subscripts LL and LH represent the low frequency and high frequency respectively during the second-level decomposition.

[0056] Time series data usually shows periodic changes, which are mainly affected by factors such as operation mode, seasonal changes, weather conditions, production cycle, and maintenance activities. The autocorrelation coefficient in the stochastic process can be used to evaluate the similarity between the sequence s t and its τ-delay s t-τ and can also be regarded as the confidence level with a delay length of τ. The larger the , the more periodic the sequence s t is on τ. By using the autocorrelation coefficient, the estimated period length can be obtained, and then the Roll operation can be used for sequence alignment and then delay information aggregation. Among them, Roll(s,τ) represents rolling the sequence s by τ steps. For a certain sequence s, its query matrix Q, key matrix K, and value matrix V are obtained through a certain transformation (such as fully connected or linear transformation), and the largest κ autocorrelation coefficients are taken for Softmax normalization processing, where τ is the delay length, to obtain τ1,…,τ κ , and then the autocorrelation attention is obtained, that is:

[0057]

[0058] After passing the low-frequency sequence of the first-level fusion and the high-frequency sequence of the first-level fusion of the i-th sensor through the autocorrelation attention module, residual connection is performed. The result of the residual connection is merged after passing through the one-dimensional convolutional network 1DCNN to obtain the first-level time series feature map of the i-th sensor. The module composed of the above operations is called the autocorrelation fusion module (ACF):

[0059]

[0060] where d is the number of kernels of the one-dimensional convolutional network 1DCNN. For and In a similar manner, the second-level time series feature map of the i-th sensor can be obtained and collectively referred to as the time series feature map.

[0061] After obtaining the multi-frequency components using the discrete wavelet transform, these components can be used to model the multi-scale time dependence to detect time anomalies. In addition to time anomalies, it is also necessary to handle the abnormal changes in the relationships between variables in the multi-dimensional time series, that is, sensor anomalies. For the anomalies caused by the changes in the correlations between variables, the present invention introduces a dynamic graph module to capture the correlations of the decomposed multi-scale frequency components.

[0062] Model the spatial dependence relationship between variables at each scale as a dynamic graph network, define the variable features as nodes, define the relationships between variables as edges, and use a graph convolutional network to aggregate the features between nodes. In system anomalies, it often occurs that one abnormal variable will cause anomalies in other variables, that is, the influence of the anomaly is unidirectional. Therefore, it is necessary to construct an asymmetric adjacency matrix. Input Z i into the graph convolutional network GCN to aggregate the feature maps of different variables and obtain the aggregation result V i , that is

[0063] V i = GCN(Z i ).

[0064] If a one-dimensional convolutional network is used to transform the aggregation result at the second level, the second-level low-frequency sequence corresponding to is obtained as well as the second-level high-frequency sequence corresponding to

[0065]

[0066] Similarly, a one-dimensional convolutional network is used to transform the aggregation result V i at the first level to obtain the first-level high-frequency sequence corresponding to

[0067]

[0068] and the first-level low-frequency sequence is obtained by the inverse multi-wavelet basis function wavelet transform mbIDWT of , that is:

[0069]

[0070] Furthermore, by​​ and Perform multi-wavelet basis function inverse wavelet transform mbIDWT to obtain the reconstructed data

[0071]

[0072] Obviously, Reconstructed time series data for each sensor is obtained by averaging the columns Composition of complete reconstructed multidimensional time series data

[0073] 3. Define the loss function for model training:

[0074] In the reconstruction process, the inverse wavelet transform recursively synthesizes the original time series, in which the reconstruction errors of slight anomalies will be accumulated and amplified. Therefore, the following loss function is defined:

[0075]

[0076] Among them, M X is a mask matrix, corresponding to X. If a position in X is missing and filled with zero, the element at the corresponding position in M is 0, otherwise it is 1. Similarly, the mask matrix M H ,M LH ,M LL Respectively correspond; Indicates the multiplication of corresponding elements; λ X ,λ H ,λ LH ,λ LL Represents the weighting coefficient.

[0077] 4. Abnormal diagnosis and missing completion:

[0078] Subtract the reconstructed time series data from the original time series data to obtain:

[0079]

[0080] When Δ exceeds a certain set threshold, the data for that period is considered to be an outlier, which is removed and padded with zeros, and the mask matrix is updated. The model is trained again until no new outliers are found or a specific number of cycles is reached. All outliers have been removed and missing values have been completed.

[0081] Among them, X is the multidimensional time series data used in the training phase, and X0 is the multidimensional time series data obtained in actual production.

[0082] The present invention also provides a device for diagnosing and complementing multi-dimensional time series data. The device in the present invention corresponds to the method, and is applicable to the specific technical solutions of the method, and is also applicable to the system.

[0083] Those skilled in the art should understand that the embodiments of the present invention can be provided as a method, a system, or a computer program product. Therefore, the present invention can take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware aspects. Moreover, the present invention can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) that contain computer-usable program code.

[0084] For those skilled in the art, it is obvious that the present invention is not limited to the details of the above exemplary embodiments, and can be implemented in other specific forms without departing from the spirit or basic characteristics of the present invention. Therefore, from any point of view, the embodiments should be regarded as exemplary and non-limiting. The scope of the present invention is defined by the appended claims rather than the above description. Therefore, all changes falling within the meaning and scope of the equivalent elements of the claims are intended to be included in the present invention, and any reference signs in the claims should not be regarded as limiting the claims involved.

[0085] In addition, it should be understood that although this specification is described according to embodiments, not every embodiment only contains an independent technical solution. This narrative way of the specification is only for clarity. Those skilled in the art should regard the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.

Claims

1. A method for diagnosing and complementing multi-dimensional time series data, characterized in that, Specifically, it includes the following steps: Step 1, collect the sensor time series data generated by n sensors on the production equipment to obtain multi-dimensional time series data where x i represents the sensor time series data generated by the i-th sensor, 1 ≤ i ≤ n; is the w-th sensor data in x i where w is the length of the sensor time series data of each sensor; Step 2, construct a neural network model: For each x in X i perform the first-level discrete wavelet transform using multi-wavelet basis functions to obtain the low-frequency component of the first-level transform of x i and the high-frequency component of the first-level transform Input them into a one-dimensional convolutional network and a fully connected network respectively to obtain the first-level fused low-frequency sequence and the first-level fused high-frequency sequence Perform the second-level discrete wavelet transform using multi-wavelet basis functions on to obtain the low-frequency component of the second-level transform and the high-frequency component of the second-level transform Input them into a one-dimensional convolutional network and a fully connected network respectively to obtain the second-level fused low-frequency sequence and the second-level fused high-frequency sequence Input and into a self-correlation fusion module composed of self-correlation attention, residual connection, and a one-dimensional convolutional network to obtain the first-level time series feature map of the i-th sensor Input and into the self-correlation fusion module to obtain the second-level time series feature map of the i-th sensor Input the aggregation result obtained after inputting into the graph convolutional network into a one-dimensional convolutional network to obtain a first-level high-frequency sequence Input the aggregation result obtained after inputting into the graph convolutional network into a one-dimensional convolutional network to obtain a second-level low-frequency sequence and a second-level high-frequency sequence Perform inverse wavelet transform of multi-wavelet basis functions on and to obtain a first-level low-frequency sequence By performing inverse wavelet transform of multi-wavelet basis functions on and to obtain the reconstructed data By taking the average of column by column to obtain the reconstructed time series data of the i-th sensor Furthermore, obtain the reconstructed multi-dimensional time series data Step 3, define the loss function of the neural network model and perform training; the loss function is as follows: Among them, M X is the mask matrix corresponding to X; if X is filled with zero due to missing at a specific position, the element of M at the specific position is 0, and in other cases, the element of M at the specific position is 1; M H , M LH , M LL are respectively the mask matrices corresponding to ; represents the multiplication of corresponding elements; λ X , λ H , λ LH , λ LL represent the weighting coefficients, and ‖·‖2 represents the second norm; Step 4, anomaly diagnosis and code completion, specifically including the following steps: Step S41: Input the multi-dimensional time series data X0 obtained in actual production into the trained neural network model to obtain the reconstructed multi-dimensional time series data Calculate the observation error Δ: If the observation error Δ exceeds the set threshold, the corresponding sensor data in X0 is removed as an outlier and filled with zero, and the mask matrix is updated, and the neural network model is retrained; the updated multi-dimensional time series data X0 is input into the retrained neural network model; Step S42, loop through step S41 until no new outliers can be found or a specific number of loops is reached. Finally, all outlier sensor data in the multi-dimensional time series data X0 has been removed and missing value imputation has been achieved. All outlier sensor data in the multi-dimensional time series data X0 has been removed and missing value imputation has been achieved.

2. The multi-dimensional time series data diagnosis and completion method according to claim 1, characterized in that The wavelet basis functions used in the multi-wavelet basis function discrete wavelet transform include two or more wavelet basis functions among Haar wavelet, Daubechies wavelet, Mexican Hat wavelet, Morlet wavelet, and Meyer wavelet.

3. The method for diagnosing and complementing multi-dimensional time series data according to claim 1, characterized in that In step two, for each x in X i perform the first-level discrete wavelet transform with multi-wavelet basis functions to obtain the first-level transformed low-frequency component i and the first-level transformed high-frequency component corresponding to x Input into a one-dimensional convolutional network and a fully connected network respectively to obtain the first-level fused low-frequency sequence and the first-level fused high-frequency sequence Specifically, it includes: The time series data x of the i-th sensor i After discrete wavelet transform with multi-wavelet basis functions, the low-frequency component of the first-level transform is obtained and the high-frequency component of the first-level transform Among them, mbDWT(·) represents the discrete wavelet transform of multi-wavelet basis functions, and k is the total number of wavelet basis functions in the discrete wavelet transform of multi-wavelet basis functions; represents x i The low-frequency component of the discrete wavelet transform under the j-th wavelet basis function, where 1 ≤ j ≤ k; Pair are fused to obtain the first-level fused low-frequency sequence Specifically, a multi-core one-dimensional convolutional network 1DCNN is used to transform the discrete wavelet transform low-frequency components corresponding to each wavelet basis function The transformation results are input into a fully connected network FC for fusion to obtain the first-level fused low-frequency sequence Input the first-level transformed high-frequency components into the one-dimensional convolutional network 1DCNN and the fully connected network FC of multiple cores to obtain the first-level fused high-frequency sequence 4. The multi-dimensional time series data diagnosis and completion method according to claim 1, characterized in that In step two, the and are input into a self - correlation fusion module composed of self - correlation attention, residual connection, and one - dimensional convolutional network to obtain the first - level time - series feature map of the i - th sensor The and are input into the self - correlation fusion module to obtain the second - level time - series feature map of the i - th sensor Specifically, it includes: The first-level fusion low-frequency sequence of the i-th sensor and the first-level fusion high-frequency sequence are respectively input into the self-correlation attention AC and residual connection is performed. The obtained results are input into the one-dimensional convolutional network 1DCNN and then merged to obtain the first-level time series feature map of the i-th sensor Where d is the number of kernels in the one-dimensional convolutional network, and || represents merging; Input the second-level fusion low-frequency sequence and the second-level fusion high-frequency sequence into the self-correlation attention AC respectively and perform residual connection. Input the obtained results into the one-dimensional convolutional network 1DCNN and then merge them to obtain the second-level time series feature map of the i-th sensor 5. A multi-dimensional time series data diagnosis and completion device, characterized in that, It includes: Data collection module, which collects the sensor time series data generated by n sensors on the production equipment to obtain multi-dimensional time series data Among them, x i represents the sensor time series data generated by the i-th sensor, where 1 ≤ i ≤ n; is the w-th sensor data in x i where w is the length of the sensor time series data of each sensor; The neural network model construction module performs the first-level discrete wavelet transform of each \(x\) in \(X\) using multi-wavelet basis functions to obtain the first-level transformed low-frequency component of \(x\) i and the first-level transformed high-frequency component of \(x\) i The first-level transformed low-frequency component and the first-level transformed high-frequency component are respectively input into a one-dimensional convolutional network and a fully connected network to obtain the first-level fused low-frequency sequence and the first-level fused high-frequency sequence The first-level fused high-frequency sequence Performs the second-level discrete wavelet transform of the first-level fused high-frequency sequence using multi-wavelet basis functions to obtain the second-level transformed low-frequency component and the second-level transformed high-frequency component The second-level transformed high-frequency component are respectively input into a one-dimensional convolutional network and a fully connected network to obtain the second-level fused low-frequency sequence and the second-level fused high-frequency sequence The second-level fused high-frequency sequence and the second-level fused low-frequency sequence are input into a self-correlation fusion module composed of self-correlation attention, residual connection, and one-dimensional convolutional network to obtain the first-level time series feature map of the \(i\)-th sensor The first-level time series feature map of the \(i\)-th sensor and the second-level fused high-frequency sequence are input into the self-correlation fusion module to obtain the second-level time series feature map of the \(i\)-th sensor The second-level time series feature map of the \(i\)-th sensor Input the aggregation result obtained after inputting into the graph convolutional network into a one-dimensional convolutional network to obtain a first-level high-frequency sequence Input the aggregation result obtained after inputting into the graph convolutional network into a one-dimensional convolutional network to obtain a second-level low-frequency sequence and a second-level high-frequency sequence Perform the inverse wavelet transform of multi-wavelet basis functions on and to obtain a first-level low-frequency sequence Reconstructed data is obtained by performing the inverse wavelet transform of multi-wavelet basis functions on and The reconstructed time series data of the i-th sensor is obtained by taking the average of column by column Furthermore, the reconstructed multi-dimensional time series data is obtained The training module defines the loss function of the neural network model and conducts training; the loss function is as follows: where M X is the mask matrix corresponding to X; if X is filled with zero due to missing at a specific position, the element of M at the said specific position is 0; M H , M LH , M LL are the mask matrices corresponding to respectively; represents element-wise multiplication; λ X , λ H , λ LH , λ LL represent the weighting coefficients, and ‖·‖2 represents the second norm; Anomaly diagnosis and code completion module, which inputs the multi-dimensional time series data X0 obtained in actual production into the trained neural network model to obtain the reconstructed multi-dimensional time series data Calculate the observation error Δ: If the observation error Δ exceeds the set threshold, the corresponding sensor data in X0 is removed as an outlier and filled with zero, and the mask matrix is updated, and the neural network model is retrained; the updated multi-dimensional time series data X0 is input into the retrained neural network model; the training process is looped until no new outliers can be found or a specific number of loops is reached, and finally the All abnormal sensor data in the multi-dimensional time series data X0 have been removed and missing values have been filled in 6. A multi-dimensional time series data diagnosis and completion system, including a memory and a processor. The memory stores a computer program. When the computer program is executed by the processor, the processor is caused to execute the steps of the method according to any one of claims 1 to 4.

7. A computer-readable storage medium stores a computer program. When the computer program is executed by a processor, the processor is caused to execute the steps of the method according to any one of claims 1 to 4.

Images