A complex equipment anomaly detection method based on a double-time convolution network

Abstract

The application relates to the technical field of equipment health management, in particular to a complex equipment anomaly detection method based on a double-time convolution network, which comprises the following steps: collecting historical operation data of complex equipment through a sensor as training data; obtaining a trained time sequence prediction network based on the training data; obtaining prediction deviation data from the trained time sequence prediction network and determining a prediction deviation basic threshold; obtaining a trained prediction deviation estimation network from the training data and the prediction deviation data; collecting online operation data of the complex equipment through the sensor; processing the online operation data to obtain actual prediction deviation at a future moment, comparing the actual prediction deviation with an adaptive threshold, and determining whether the operation state of the complex equipment at the future moment is abnormal; and the application can improve the prediction anomaly detection accuracy of the complex equipment.

Description

Technical Field

[0001] This invention relates to the field of equipment health management technology, specifically to a method for detecting anomalies in complex equipment based on dual-temporal convolutional networks. Background Technology

[0002] Complex industrial equipment integrates mechanical, hydraulic, electrical, magnetic, and thermal components, processing massive amounts of multidimensional operational data characterized by high volume, complexity, and interdependence. Anomaly detection plays a crucial role in monitoring the operational status of this equipment, achieving this by identifying data deviations that could indicate potential signal failures. Fast and accurate anomaly detection enables timely maintenance, ensuring system reliability and efficiency. Data-driven technologies based on machine learning and deep learning have effectively replaced traditional thresholding methods based on expert experience. However, machine learning techniques such as clustering and classification often struggle with high-dimensional and complex data, limiting their applicability. In contrast, deep learning methods excel at uncovering complex patterns in large, high-dimensional data. Currently, a representative example is prediction-based anomaly detection methods, which detect anomalies by evaluating the deviation between predicted and actual data points.

[0003] In current prediction-based anomaly detection methods, a threshold for prediction bias is typically set based on the overall prediction error of the prediction model on known normal training data. Existing methods usually assume a uniform tolerance for prediction bias across all inputs, i.e., setting a fixed threshold for prediction bias. However, the training objective of a prediction model is to minimize the global error. Therefore, data patterns that occur frequently in the training data will lead to better learning and more accurate predictions. Since the training data itself is unevenly distributed, the trained model will perform differently on different inputs, making it insufficient to set a prediction bias threshold for all samples based on the global prediction bias of the training data. Summary of the Invention

[0004] In view of the above problems, the present invention provides a method for detecting anomalies in complex equipment based on dual temporal convolutional networks, which solves the technical problem of low accuracy in predicting and detecting anomalies in complex equipment in the prior art.

[0005] This invention provides a method for detecting anomalies in complex equipment based on dual-temporal convolutional networks, comprising the following steps:

[0006] Step S1: Install sensors on the complex equipment and collect historical operating data of the complex equipment as training data. The training data includes multiple historical samples arranged in chronological order, and the historical samples include multiple operating parameter values ​​of the complex equipment.

[0007] Step S2: Determine the network structure of the time series prediction network, and train the time series prediction network based on the training data to obtain the trained time series prediction network.

[0008] Step S3: The trained time series prediction network processes the training data to obtain prediction bias data, and determines the basic threshold of prediction bias based on the prediction bias data.

[0009] Step S4: Determine the network structure of the prediction bias estimation network, and train the prediction bias estimation network based on the training data and the prediction bias data to obtain the trained prediction bias estimation network.

[0010] Step S5: Collect online operating data of complex equipment through sensors; the online operating data includes online samples at the current moment and online samples from multiple moments before the current moment, and the online samples include multiple operating parameter values ​​of the complex equipment;

[0011] The trained time series prediction network and the trained prediction bias estimation network are used to process the online running data to obtain the actual prediction bias at future times, and an adaptive threshold is determined by the prediction bias base threshold.

[0012] The actual prediction deviation at future moments is compared with the adaptive threshold to determine whether the operating state of the complex equipment at future moments is abnormal.

[0013] Preferably, step S2 specifically includes:

[0014] Step S2-1: Determine the network structure of the time series prediction network as follows: take multiple historical samples arranged in chronological order in the training data as multidimensional time series data as the input layer, pass through 3 stacked TCN blocks and a fully connected layer in sequence, and finally pass through the output layer to obtain the predicted values ​​of the running parameters output by the time series prediction network.

[0015] Step S2-2: Using the operating parameters of the past 50 historical samples in the training data as network input and the operating parameters of the next historical sample as the supervision target, determine the loss function of the time series prediction network. P To minimize loss P The time series prediction network is trained using the optimization objective to obtain the trained time series prediction network.

[0016] Preferably, the loss function of the time series prediction network is loss P The expression is:

[0017]

[0018] Among them, y i , are the supervised target value and predicted value of the running parameters of the i-th historical sample, respectively, and N is the number of training samples.

[0019] Preferably, step S3 specifically includes:

[0020] Step S3-1: Calculate the prediction bias for each historical sample in the training data using the trained time series prediction network. This prediction bias data is calculated as follows:

[0021]

[0022] Where, r i This represents the prediction bias of the i-th historical sample in the training data.

[0023] Step S3-2: Based on the prediction bias data of the training data, set a basic threshold for the prediction bias using the extreme value theory method. The expression is:

[0024]

[0025] Among them, z q The prediction bias threshold is defined by t, which represents the preset peak value, q, and N, which represents the number of training samples. t γ is the number of samples in the sample that exceed the peak value t, and γ and σ are the shape and scale parameters of the Pareto distribution, respectively. These are the estimated values ​​of γ and σ, respectively.

[0026] Preferably, step S4 specifically includes:

[0027] Step S4-1: Determine the network structure of the prediction bias estimation network as follows: take multiple historical samples arranged in chronological order in the training data as multidimensional time series data as the input layer, pass through 3 stacked TCN blocks and fully connected layers in sequence, and finally pass through the output layer to obtain the prediction bias estimate value output by the prediction bias estimation network.

[0028] Step S4-2: Using the running parameters of the past 50 historical samples in the training data as network input, and taking the prediction deviation corresponding to the next historical sample in the prediction deviation data as the supervision target, determine the loss function of the prediction deviation estimation network. R To minimize loss R The time series prediction network is trained using the optimization objective to obtain the trained time series prediction network.

[0029] Preferably, the loss function of the prediction bias estimation network is...R The expression is:

[0030]

[0031] Where, r i Let the prediction bias of the i-th historical sample be used as the supervision target. This represents the estimated prediction bias for the i-th historical sample.

[0032] Preferably, step S5 specifically includes:

[0033] Step S5-1: Collect online operation data of complex equipment through sensors. The online operation data includes online samples at the current moment and online samples from multiple moments before the current moment. The online samples include multiple operating parameter values ​​of the complex equipment.

[0034] Step S5-2: Process the online running data using the trained time series prediction network to obtain predicted values ​​of running parameters at future times; process the online running data using the trained prediction bias estimation network to obtain estimated values ​​of prediction bias at future times.

[0035] Step S5-3-1: Obtain the actual prediction deviation for the future time based on the predicted values ​​of the operating parameters at the future time;

[0036] Step S5-3-2: Determine the adaptive threshold based on the estimated prediction deviation at the future time, the actual prediction deviation at the future time, and the basic prediction deviation threshold.

[0037] Step S5-3-3: Compare the actual prediction deviation at the future time with the adaptive threshold to determine whether the operating status of the complex equipment at the future time is abnormal.

[0038] Preferably, step S5-3-1 specifically includes:

[0039] The absolute value of the difference between the predicted value of the operating parameters at the future time and the actual value of the operating parameters at the future time is calculated as the actual prediction deviation at the future time.

[0040] In step S5-3-2, the adaptive threshold th ad + The calculation expression is:

[0041]

[0042] in, This is an estimate of the prediction bias at future times. Let represent the prediction bias of the i-th historical sample in the training data, where N is the number of training samples and r represents the actual prediction bias at future times.

[0043] The specific steps of step S5-3-3 include: if the actual prediction deviation exceeds the adaptive threshold, it is detected as an abnormal state; otherwise, it is normal.

[0044] Preferably, the TCN block comprises a causal convolutional layer, a dilated convolutional layer, and a residual connection output layer connected in sequence, wherein,

[0045] The calculation expression for a causal convolutional layer is:

[0046]

[0047] Where, x T F(x) represents the T-th value of the input time series of the causal convolutional layer. T ) represents x T The corresponding causal convolution value, f k This represents the k-th value in the convolution kernel, where K represents the kernel length, and x... T-K+k This represents the (T-K+k)th value of the input time series of the causal convolutional layer;

[0048] The calculation expression for a dilated convolutional layer is:

[0049]

[0050] Where Fd(x) T ) represents x T The corresponding dilated convolution value, x T-(K-k)d This represents the T-(Kk)d-th value of the input time series of the causal convolutional layer, where d represents the dilation coefficient and c is the number of the network layers of the dilated convolutional layer.

[0051] The calculation expression for the residual connection output layer is:

[0052] F(x T ) = F d (x T )+x T

[0053] Where F(x) T ) represents x T The corresponding residual connection output value.

[0054] Compared with the prior art, the present invention has at least the following beneficial effects:

[0055] (1) This invention effectively improves the real-time performance and accuracy of anomaly detection in complex equipment by introducing a dual temporal convolutional network (TCN). First, compared with other deep learning algorithms, TCN has higher computational efficiency and can process large amounts of online running data in real time. This ensures that abnormal situations can be detected and responded to in a timely manner during the operation of complex equipment, thus guaranteeing the safe and stable operation of the system.

[0056] (2) This invention introduces an adaptive threshold mechanism, overcoming the limitations of static thresholds. By analyzing the differences in prediction bias between different inputs, the adaptive threshold can be dynamically adjusted, effectively managing the tolerance for prediction bias caused by uneven distribution of training data. This improves the accuracy of anomaly detection, enabling the system to maintain high detection performance even in the face of changing operating environments.

[0057] (3) This invention combines a P-network for time series prediction and an R-network for bias estimation, uses extreme value theory (EVT) to set a basic threshold, and performs adaptive adjustment through the R-network to achieve high-precision anomaly detection. It can promptly identify early potential faults in complex equipment, and also provides strong protection for the safety and stability of the equipment, improving the reliability and operating efficiency of the system. Attached Figure Description

[0058] The accompanying drawings are for illustrative purposes only and are not intended to limit the scope of the invention.

[0059] Figure 1 A flowchart of the complex equipment anomaly detection method based on dual temporal convolutional networks provided by the present invention;

[0060] Figure 2 This is a schematic diagram of the structure of the complex equipment anomaly detection method based on dual-temporal convolutional networks provided by the present invention.

[0061] Figure 3 This is a schematic diagram of the computation of a temporal convolutional network.

[0062] Figure 4 The structure diagram of the sequence prediction network and prediction bias estimation network provided by the present invention.

[0063] Figure 5 This is a schematic diagram illustrating the construction of training data for the sequence prediction network and prediction bias estimation network provided by the present invention.

[0064] Figure 6 This is a schematic diagram of the basic threshold based on EVT in this invention.

[0065] Figure 7 This is a schematic diagram of an abnormality detection threshold for a certain equipment provided by the present invention.

[0066] Figure 8This is a schematic diagram of the abnormal detection results of a certain equipment provided by the present invention. Detailed Implementation

[0067] To better understand the above-described objectives, features, and advantages of the present invention, the invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be noted that, unless otherwise specified, the embodiments of the present invention and the features thereof can be combined with each other. Furthermore, the present invention can be implemented in other ways different from those described herein; therefore, the scope of protection of the present invention is not limited to the specific embodiments disclosed below.

[0068] To illustrate the effectiveness of the method proposed in this invention, the following detailed description of the above technical solution is provided through a specific embodiment, such as... Figure 1 , Figure 2 As shown, a method for detecting anomalies in complex equipment based on dual-temporal convolutional networks is disclosed. The specific implementation steps are as follows:

[0069] Step S1: Install sensors on the complex equipment and collect historical operating data of the complex equipment as training data. The training data includes multiple historical samples arranged in chronological order, and the historical samples include multiple operating parameter values ​​of the complex equipment.

[0070] This invention uses a large amount of data on normal states accumulated during the historical operation of complex equipment as training data for a neural network. Specifically, different types of sensors can be installed on the complex equipment. These sensors can continuously collect various operating parameters during the equipment's historical operation, including physical quantities such as temperature, pressure, vibration, current, and rotational speed. The sampling frequency for collecting these operating parameters can be set according to actual operating conditions, such as sampling once per second or once per minute.

[0071] In some embodiments, the total time span for collecting operating parameters of complex equipment can be no less than 3 months to ensure that enough operating condition data are collected. After preprocessing, including outlier removal and data normalization, the collected data is stored in a database as training data for subsequent neural network training.

[0072] In some embodiments, the training data of the present invention is stored and processed in the form of time-series data. Specifically, the training data includes multiple historical samples arranged in chronological order. Each historical sample includes a timestamp and corresponding measurement values ​​from different sensors. The data structure can be: the first column is the timestamp, and subsequent columns are the measurement values ​​from different sensors, all represented as floating-point numbers. For example, for the case of n sensors, the format of each historical sample record is: [timestamp, value_1, value_2, ..., value_n]. Wherein, value_1 to value_n represent the measurement values ​​of multiple different operating parameters of different sensors at that timestamp.

[0073] Step S2: Determine the network structure of the time series prediction network, and train the time series prediction network based on the training data to obtain the trained time series prediction network.

[0074] The time series prediction network of this invention is based on a Temporal Convolutional Network (TCN). For example... Figure 3 The diagram illustrates the computation of a temporal convolutional network (TCN). The TCN network requires pre-setting parameters including the total number of layers L and the kernel size. The computation of each layer consists of three components: causal convolution, dilated convolution, and residual connections, as detailed below:

[0075] (1) Causal convolution

[0076] like Figure 3 As shown, for each layer's input, the time series data X = [x1, x2, ..., x...] is of length T. T The convolution kernel of length K is F = [f1, f2, ..., f...]. K For each time step t, convolution is performed using only data from time step t and times prior to it. The expression for calculating the causal convolution value is:

[0077]

[0078] Where, x T F represents the T-th value of the input time series of the causal convolutional layer. c (x T ) represents x T The corresponding causal convolution value, f k This represents the k-th value in the convolution kernel, where K represents the kernel length, and x... T-K+k This represents the (T-K+k)th value of the input time series of the causal convolutional layer.

[0079] (2) Dilated convolution

[0080] TCN networks improve upon the aforementioned causal networks, such as... Figure 2 As shown, for each layer of input time series data, an expansion coefficient d = 2 is used. c The network uses dilated convolutions, where c is the number of network layers. Specifically, the dilated convolution in the c-th layer inserts d-1 holes between adjacent inputs, allowing the convolutional kernel to cover a wider range of historical data. For example, when c=1 and d=2, the kernel samples every other data point in the input sequence; when c=2 and d=4, it samples every three data points. This approach significantly increases the network's receptive field, enabling it to capture longer-term temporal dependencies.

[0081] Therefore, the expression for the dilated causal convolution of the c-th layer is:

[0082]

[0083] d=2 c

[0084] Where, x T F represents the T-th value of the input time series of the dilated convolutional layer. d (x T ) represents x T The corresponding dilated convolution value, x T-(K-k)d This represents the T-(Kk)d-th value of the input time series of the causal convolutional layer, where d represents the dilation coefficient and c is the number of the dilated convolutional layers.

[0085] (3) Residual connections

[0086] The residual connections in TCN directly pass the original input to the output via shortcut connections, performing element-wise addition with the transformed features. This approach effectively alleviates the vanishing gradient problem during deep network training, improving the network's training performance. Specifically, the final output is obtained by adding the input time-series values ​​of the causal convolutional layer to the features processed by dilated convolution, expressed as:

[0087] F(x T ) = F d (x T )+x T

[0088] Where F(x) T ) represents x T The corresponding residual connection output value.

[0089] The network composed of causal convolution, dilated convolution, and residual connections is also called a TCN block. The network structure of the time series prediction network (also known as the P network) of this invention is as follows: Figure 4As shown, multiple historical samples arranged in chronological order in the training data are used as multidimensional time series data as the input layer. The data is then passed through three stacked TCN blocks and a fully connected layer, finally yielding the predicted values ​​of the running parameters output by the time series prediction network.

[0090] The steps for training the time series prediction network include: using the operating parameters of the past 50 historical samples in the training data as network input, and the operating parameters of the next historical sample as the supervision target. Determining the loss function for training the time series prediction network, specifically the P-network loss function. P The root mean square error of the predicted value is expressed as:

[0091]

[0092] Among them, y i , are the supervised target value and predicted value of the running parameters of the i-th historical sample, respectively, and N is the number of training samples.

[0093] In some embodiments, error backpropagation algorithms such as gradient descent can be used to minimize the loss function of the P network. P The time series prediction network is trained using this as the optimization objective, resulting in a fully trained time series prediction network. While there are various mature neural network training and optimization methods in the prior art, this invention does not limit the specific optimization method used for neural network training.

[0094] Step S3: The trained time series prediction network processes the training data to obtain prediction bias data, and determines the basic threshold of prediction bias based on the prediction bias data.

[0095] In this step, such as Figure 5 As shown, the prediction bias of each historical sample in the training data is calculated using a trained time series prediction network. The prediction bias r of the i-th historical sample in the training data is... i The expression is:

[0096]

[0097] like Figure 6 As shown, based on the prediction bias data of the training data, a basic threshold for the prediction bias is set using extreme value theory. Extreme value theory (EVT) posits that extreme events of different things follow the same distribution; that is, the portion of an extreme value exceeding a threshold follows a generalized Pareto distribution (GPD), which can be expressed as follows:

[0098]

[0099] Where Z represents a prediction bias value, z represents the prediction bias random variable, and t represents a preset threshold. This represents the probability distribution of the extreme values ​​of the prediction bias exceeding t. Let τ represent the conditional probability that the excess Zt is greater than the prediction bias random variable z, given that Z is greater than t. τ represents the peak value. This indicates that when t is set to a peak value, it follows a Pareto distribution with parameters γ and σ, where γ is the shape parameter and σ is the scale parameter.

[0100] Estimates of parameters γ and σ The prediction bias baseline threshold z can be obtained using the Method of Moments (MOM). q The calculation expression is as follows:

[0101]

[0102] Where q is the target probability parameter, the value of which is set manually, usually 0.01%, and N is the number of training samples. t It is the number of samples in the sample that exceed the peak value t. These are the estimated values ​​of γ and σ, respectively.

[0103] Step S4: Determine the network structure of the prediction bias estimation network, and train the prediction bias estimation network based on the training data and the prediction bias data to obtain the trained prediction bias estimation network.

[0104] like Figure 4 As shown, the prediction bias estimation network (also known as the R network) of the present invention has the same network structure as the time series prediction network (also known as the P network).

[0105] Multiple historical samples arranged in chronological order in the training data are used as input to multidimensional time series data. The data passes through an input layer, three stacked TCN blocks, and a fully connected layer, and finally the prediction bias estimate of the time series prediction network output is obtained.

[0106] The steps for training the prediction bias estimation network include: using the running parameter data of the past 50 historical samples in the training data as network input, and using the prediction bias corresponding to the next historical sample in the prediction bias data as the supervision target. Determine the loss function for training the prediction bias estimation network, specifically the R network loss function. R The root mean square error of the prediction bias is expressed as:

[0107]

[0108] Where, r i Let the prediction bias of the i-th historical sample be used as the supervision target. This represents the estimated prediction bias for the i-th historical sample.

[0109] In some embodiments, similarly, error backpropagation algorithms such as gradient descent can be employed to minimize the loss function of the R network. R The prediction bias estimation network is trained using this as the optimization objective, resulting in a trained prediction bias estimation network. While there are various mature neural network training optimization methods in the prior art, this invention does not limit the specific optimization method used for neural network training.

[0110] Step S5-1: Collect online operating data of complex equipment through sensors. The online operating data includes online samples at the current moment and online samples from multiple moments prior to the current moment. The online samples include multiple operating parameter values ​​of the complex equipment.

[0111] After the P-network and R-network are trained, the online operational data of complex equipment can be monitored. Here, "online" specifically refers to the real-time nature of the data, meaning that data acquisition, transmission, and processing are all carried out in real time during equipment operation, rather than offline analysis afterward.

[0112] The method for acquiring online operational data in this invention is similar to that for acquiring historical operational data, also involving continuous collection of various operational parameters of the equipment during operation via sensors. Specifically, the operational parameter values ​​at the current moment are saved, as well as the operational parameter values ​​from multiple moments prior to the current moment. In some embodiments, the operational parameter values ​​of the latest 50 online samples prior to the current moment can be saved.

[0113] Similar to the historical samples of this invention, the online samples of this invention also include timestamps and corresponding measurements from different sensors. The data structure can be: the first column is the timestamp, and the subsequent columns are the measurements from different sensors.

[0114] Step S5-2: Use the trained time series prediction network to process the online running data to obtain the predicted values ​​of the running parameters at future times;

[0115] The trained prediction bias estimation network is used to process the online running data to obtain prediction bias estimates for future time periods.

[0116] In some embodiments, the latest 50 online samples can be input into the trained time series prediction network to obtain the predicted values ​​of the operating parameters at future times; the latest 50 online samples can be input into the trained prediction bias estimation network to obtain the estimated values ​​of the prediction bias at future times.

[0117] Step S5-3-1: Obtain the actual prediction deviation for the future time based on the predicted values ​​of the operating parameters at the future time;

[0118] Step S5-3-2: Determine the adaptive threshold based on the estimated prediction deviation at the future time, the actual prediction deviation at the future time, and the basic prediction deviation threshold.

[0119] Step S5-3-3: Compare the actual prediction deviation at the future time with the adaptive threshold to determine whether the operating status of the complex equipment at the future time is abnormal.

[0120] The adaptive threshold of this invention is dynamically and adaptively adjusted based on the prediction deviation estimate output by the R network, building upon the basic prediction deviation threshold. The unoptimized adaptive threshold th... ad The expression is:

[0121]

[0122] in, This is an estimate of the prediction bias at future times. Let N be the prediction bias of the i-th historical sample in the training data, and N be the total number of training samples.

[0123] In this step, the actual prediction bias for future moments is obtained in a manner similar to that used for historical samples. The actual prediction bias for future moments is calculated as follows: the absolute value of the difference between the predicted values ​​of the operating parameters for future moments obtained by the trained time series prediction network (P-network) based on the online operating data and the true values ​​of the operating parameters for future moments is taken as the actual prediction bias for future moments.

[0124] Besides dynamic threshold adjustment, prediction biases significantly exceeding the basic threshold can be directly classified as anomalies. To optimize the calculation of the adaptive threshold, when the actual prediction bias r satisfies r > 2*z... q If the prediction bias is considered to be significantly greater than the basic threshold, then the final adaptive threshold th ad + The expression is:

[0125]

[0126] Where r represents the actual prediction deviation at future times.

[0127] like Figure 7 The diagram shows the prediction deviation and deviation threshold during the anomaly detection of a certain equipment, including the actual prediction deviation, the estimated value of the prediction deviation output by the R network, the basic threshold based on EVT, and the adaptive dynamic threshold of the present invention.

[0128] Finally, the actual prediction deviation is compared with the adaptive threshold. If the actual prediction deviation exceeds the adaptive threshold, it is detected as an abnormal state; otherwise, it is normal.

[0129] like Figure 8 The diagram shows the results of abnormal equipment detection, demonstrating that minute anomalies in the data are accurately detected.

[0130] This invention provides a method for anomaly detection in complex equipment based on dual temporal convolutional networks, which effectively overcomes the shortcomings of existing deep learning-based anomaly detection methods and further improves the accuracy of anomaly detection. A computationally efficient TCN network is used to establish a P network for time series prediction and an R network for bias estimation, ensuring real-time anomaly detection. This invention introduces an adaptive threshold mechanism that considers the difference in prediction bias between different inputs. By combining a basic threshold based on EVT with the adaptive adjustment of the R network, the accuracy of anomaly detection is improved, thereby enhancing the safety and stability of complex equipment operation.

[0131] While the specific embodiments of the present invention depict actions or steps in a particular order, this should be understood as requiring such actions or steps to be performed in the shown specific order or sequential order, or requiring all illustrated actions or steps to be performed to achieve the desired result. In certain environments, multitasking and parallel processing may be advantageous. Similarly, although several specific implementation details are included in the above discussion, these should not be construed as limiting the scope of this disclosure. Certain features described in the context of individual embodiments may also be implemented in combination in a single implementation. Conversely, various features described in the context of a single implementation may also be implemented individually or in any suitable sub-combination in multiple implementations. The above descriptions are merely preferred embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention.

[0132] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for detecting anomalies in complex equipment based on dual-temporal convolutional networks, characterized in that, Includes the following steps: Step S1: Install sensors on the complex equipment and collect historical operating data of the complex equipment as training data. The training data includes multiple historical samples arranged in chronological order, and the historical samples include multiple operating parameter values ​​of the complex equipment. Step S2: Determine the network structure of the time series prediction network, and train the time series prediction network based on the training data to obtain the trained time series prediction network. Step S3: The trained time series prediction network processes the training data to obtain prediction bias data, and determines the basic threshold of prediction bias based on the prediction bias data. Step S4: Determine the network structure of the prediction bias estimation network, and train the prediction bias estimation network based on the training data and the prediction bias data to obtain the trained prediction bias estimation network. Step S5: Collect online operating data of complex equipment through sensors; the online operating data includes online samples at the current moment and online samples from multiple moments before the current moment, and the online samples include multiple operating parameter values ​​of the complex equipment; The trained time series prediction network and the trained prediction bias estimation network are used to process the online running data to obtain the actual prediction bias at future times, and an adaptive threshold is determined by the prediction bias base threshold. The actual prediction deviation at future moments is compared with the adaptive threshold to determine whether the operating state of the complex equipment at future moments is abnormal.

2. The method for detecting anomalies in complex equipment based on dual-temporal convolutional networks according to claim 1, characterized in that, Step S2 specifically includes: Step S2-1: Determine the network structure of the time series prediction network as follows: take multiple historical samples arranged in chronological order in the training data as multidimensional time series data as the input layer, pass through 3 stacked TCN blocks and a fully connected layer in sequence, and finally pass through the output layer to obtain the predicted values ​​of the running parameters output by the time series prediction network. Step S2-2: Using the operating parameters of the past 50 historical samples in the training data as network input and the operating parameters of the next historical sample as the supervision target, determine the loss function of the time series prediction network. P To minimize loss P The time series prediction network is trained using the optimization objective to obtain the trained time series prediction network.

3. The method for detecting anomalies in complex equipment based on dual-temporal convolutional networks according to claim 2, characterized in that, The loss function of the time series prediction network P The expression is: Among them, y i , are the supervised target value and predicted value of the running parameters of the i-th historical sample, respectively, and N is the number of training samples.

4. The method for detecting anomalies in complex equipment based on dual-temporal convolutional networks according to claim 3, characterized in that, Step S3 specifically includes: Step S3-1: Calculate the prediction bias for each historical sample in the training data using the trained time series prediction network. This prediction bias data is calculated as follows: Where, r i This represents the prediction bias of the i-th historical sample in the training data; Step S3-2: Based on the prediction bias data of the training data, set a basic threshold for the prediction bias using the extreme value theory method. The expression is: Among them, z q The prediction bias threshold is defined by t, which represents the preset peak value, q, and N, which represents the number of training samples. t γ is the number of samples in the sample that exceed the peak value t, and γ and σ are the shape and scale parameters of the Pareto distribution, respectively. These are the estimated values ​​of γ and σ, respectively.

5. The method for detecting anomalies in complex equipment based on dual-temporal convolutional networks according to claim 4, characterized in that, Step S4 specifically includes: Step S4-1: Determine the network structure of the prediction bias estimation network as follows: take multiple historical samples arranged in chronological order in the training data as multidimensional time series data as the input layer, pass through 3 stacked TCN blocks and fully connected layers in sequence, and finally pass through the output layer to obtain the prediction bias estimate value output by the prediction bias estimation network. Step S4-2: Using the running parameters of the past 50 historical samples in the training data as network input, and taking the prediction deviation corresponding to the next historical sample in the prediction deviation data as the supervision target, determine the loss function of the prediction deviation estimation network. R To minimize loss R The time series prediction network is trained using the optimization objective to obtain the trained time series prediction network.

6. The method for detecting anomalies in complex equipment based on dual-temporal convolutional networks according to claim 5, characterized in that, The loss function of the prediction bias estimation network R The expression is: Where, r i Let the prediction bias of the i-th historical sample be used as the supervision target. This represents the estimated prediction bias for the i-th historical sample.

7. The method for detecting anomalies in complex equipment based on dual-temporal convolutional networks according to claim 6, characterized in that, Step S5 specifically includes: Step S5-1: Collect online operation data of complex equipment through sensors. The online operation data includes online samples at the current moment and online samples from multiple moments before the current moment. The online samples include multiple operating parameter values ​​of the complex equipment. Step S5-2: Process the online running data using the trained time series prediction network to obtain predicted values ​​of running parameters at future times; process the online running data using the trained prediction bias estimation network to obtain estimated values ​​of prediction bias at future times. Step S5-3-1: Obtain the actual prediction deviation for the future time based on the predicted values ​​of the operating parameters at the future time; Step S5-3-2: Determine the adaptive threshold based on the estimated prediction deviation at the future time, the actual prediction deviation at the future time, and the basic prediction deviation threshold. Step S5-3-3: Compare the actual prediction deviation at the future time with the adaptive threshold to determine whether the operating status of the complex equipment at the future time is abnormal.

8. The method for detecting anomalies in complex equipment based on dual-temporal convolutional networks according to claim 7, characterized in that, Step S5-3-1 specifically includes: The absolute value of the difference between the predicted value of the operating parameters at the future time and the actual value of the operating parameters at the future time is calculated as the actual prediction deviation at the future time. In step S5-3-2, the adaptive threshold th ad + The calculation expression is: in, This is an estimate of the prediction bias at future times. Let represent the prediction bias of the i-th historical sample in the training data, N be the number of training samples, and r represent the actual prediction bias at future time. The specific steps of step S5-3-3 include: if the actual prediction deviation exceeds the adaptive threshold, it is detected as an abnormal state; otherwise, it is normal.

9. The method for detecting anomalies in complex equipment based on dual-temporal convolutional networks according to claim 8, characterized in that, The TCN block comprises a causal convolutional layer, a dilated convolutional layer, and a residual connected output layer connected in sequence, wherein, The calculation expression for a causal convolutional layer is: Where, x T F represents the T-th value of the input time series of the causal convolutional layer. c (x T ) represents x T The corresponding causal convolution value, f k This represents the k-th value in the convolution kernel, where K represents the kernel length, and x... T-K+k This represents the (T-K+k)th value of the input time series of the causal convolutional layer; The calculation expression for a dilated convolutional layer is: d＝2 c Among them, F d (x T ) represents x T The corresponding dilated convolution value, x T-(K-k)d This represents the T-(Kk)d-th value of the input time series of the causal convolutional layer, where d represents the dilation coefficient and c is the number of the network layers of the dilated convolutional layer. The calculation expression for the residual connection output layer is: F(x T )=F d (x T )+x T Where F(x) T ) represents x T The corresponding residual connection output value.

Images