Apparatus and method for estimating a lifetime indicator of a machine

Abstract

According to various embodiments, there is provided a method for estimating a life indicator of a machine performing a given manufacturing process, comprising: generating a training time series of health indices specifying a health of a respective machine, determining a range of health indices representing different degradation levels and one or more thresholds between one or more pairs of degradation levels, the one or more thresholds including a threshold representing end of useful life, training a health index prediction model using the generated training time series of health indices as historical data to continue the health index time series into the future, and using the prediction model and the one or more thresholds to estimate a life indicator of the machine performing the manufacturing process.

Description

Technical Field

[0001] This disclosure relates to apparatus and methods for estimating machine life indicators. Background Technology

[0002] Manufacturing is one of the leading industries that extensively utilizes data-driven technologies. The development of modern manufacturing processes increasingly relies on data-driven decision-making to improve efficiency, minimize downtime, and optimize maintenance. For example, predictive maintenance (PdM), which aims to anticipate machine failures, typically by estimating the remaining useful life (RUL) of a machine tool or another lifespan metric, and recommending timely actions to machine operators and production line managers, is gaining popularity in smart manufacturing because the financial benefits of such predictive actions can be significant.

[0003] However, current PdM solutions often struggle with scalability (i.e., the possibility of applying them to different types of machines and sensors), making customized settings for various machine configurations a requirement, or they heavily rely on labeled data, which is often limited in practice. Therefore, a method for efficient and flexible RUL estimation is desirable. Summary of the Invention

[0004] According to various embodiments, a method is provided for estimating the life index of a machine performing a given manufacturing process, comprising: • Receive historical data including at least one historical data time series, said historical data time series being a time series of sensor data collected during the execution of said manufacturing process; • Apply an anomaly detector to at least one historical data time series of the historical data, and use a health index generator to generate a training time series of a health index specifying the health of the corresponding machine based on the anomaly detection results of the anomaly detector. • Determine a range of health indices representing different levels of degradation and one or more thresholds between a pair or more pairs of degradation levels, said one or more thresholds including a threshold representing the end of lifespan; • Use the generated health index training time series as training data to train the health index prediction model, so as to extend the health index time series into the future; • For the machine performing the manufacturing process, the machine's lifespan is estimated based on the time series of sensor data collected during the machine's operation, using the following methods: • By applying the anomaly detector to the time series of sensor data collected during the operation of the machine and by using the health index generator, a health index time series of the machine is generated based on the time series of sensor data collected during the operation of the machine; • Using a trained health index prediction model, extend the time series of the machine's health index into the future; and • Determine the estimated lifespan index of the machine from the point in time when the machine's ongoing health index time series reaches the threshold representing the end of its lifespan.

[0005] The method described above allows for predictive maintenance with scalability across different sensor types and machines: requiring only time-series sensor data. The number of sensors and manufacturing conditions are flexible. Scalability is achieved by tracking anomalous behavior and learning to label health degradation, dynamically detecting thresholds, and making predictions. Missing labels or failures in the data can be handled robustly: the method can operate on unlabeled data when no failure data is available. Noise injection can be performed based on the properties of the underlying data under fully unsupervised conditions to train the model by simulating failure scenarios. The method performs predictions of lifetime metrics such as RUL (Remaining Usable Lifetime), TTF (Time to Failure), half-life, or 90% lifetime by combining individual modular blocks with clearly defined individual objectives (anomaly detectors, health index generators, etc.). This improves interpretability. The modular nature of the resulting framework also allows for possible adjustments, swapping certain modules, making additions and adjustments possible and efficient (i.e., scalable).

[0006] Various examples are given below.

[0007] Example 1 is a method for estimating machine life indicators as described above.

[0008] Example 2 is based on the method described in Example 1, which includes determining a range of health indices representing different levels of degradation based on the generated health index time series by applying a machine learning cluster model.

[0009] This allows for the definition of degradation thresholds without utilizing historical time series data labeled with different degradation levels.

[0010] Example 3 is a method according to Example 1 or 2, which includes generating a historical data time series by injecting noise representing degradation into a time series of sensor data collected during the operation of a corresponding machine performing a given manufacturing process.

[0011] Therefore, if the time-series data in the historical data does not show degradation, the injected noise can be used to augment the historical data, altering the sensor data as if degradation had occurred during the operation of the corresponding machine.

[0012] Example 4 is a method according to any one of Examples 1 to 3, wherein the health index generator generates a health index by changing (e.g., reducing) an initial health index representing a healthy machine over time (i.e., for each time index of the corresponding time series) based on anomaly detection results.

[0013] The health index thus indicates that the machine was healthy at the beginning, and that its health gradually deteriorates. Each detected anomaly is considered an indication of deteriorated health.

[0014] Example 5 is the method according to Example 4, wherein the health index generator includes a first neural network that processes the anomaly detection results of the anomaly detection of the current time index and a second neural network that processes the anomaly detection results of the anomaly detection averaged over a predetermined time index window, combines the outputs of the first neural network and the second neural network, and determines the health index value of the current time index by changing (e.g., reducing) the health index value from the prior time index according to the result of the combination.

[0015] Example 6 is the method according to Example 5, which includes training a health index generator by training a first neural network and a second neural network using a loss function that includes at least one of a penalty for the deviation between a subsequent health index and a health index that is close to a health index representing maximum degradation (e.g., 0).

[0016] The first loss component ensures the stability of the health index, while the second loss component ensures that the health index only shows a high degree of degradation if there is indeed a significant degradation.

[0017] Example 7 is a method according to any one of Examples 1 to 6, wherein the anomaly detector includes an autoencoder, and the result of the anomaly detector includes the reconstruction error of the autoencoder.

[0018] This allows for efficient anomaly detection and the generation of anomaly detection results, which can then be efficiently further processed (by a health index generator to generate a health index). Furthermore, the anomaly detector can be trained using the routine training of an autoencoder (e.g., based on time-series sensor data from a health machine).

[0019] Example 8 is the method described according to Example 7, wherein the autoencoder is a temporal convolutional network autoencoder or a long short-term memory autoencoder.

[0020] This allows for efficient handling of both multivariate data (using temporal convolutional network autoencoders) and univariate data (using long short-term memory autoencoders).

[0021] Example 9 is a method for controlling a machine, comprising estimating a life index of the machine according to any one of Examples 1 to 8, and applying maintenance actions to the machine according to the estimated life index.

[0022] Example 10 is a data processing system (in particular a machine controller) configured to perform the method of any one of Examples 1 to 9.

[0023] Example 11 is a computer program that includes instructions that, when executed by a computer, cause the computer to perform the method according to any one of Examples 1 to 9.

[0024] Example 12 is a computer-readable medium including instructions that, when executed by a computer, cause the computer to perform the method according to any one of Examples 1 to 9. Attached Figure Description

[0025] In the accompanying drawings, similar reference numerals generally refer to the same parts throughout the different views. The drawings are not necessarily drawn to scale; rather, the emphasis is generally placed on illustrating the principles of the invention. In the following description, various aspects are described with reference to the following drawings, wherein: Figure 1 A winding machine is shown as an example of a manufacturing machine.

[0026] Figure 2 The illustration shows a training pipeline for a PdM (Predictive Maintenance) model, or in other words, a RUL (Remaining Useful Life) estimator, according to an embodiment.

[0027] Figure 3 The diagram illustrates the following: Figure 2 Use of the trained PdM model.

[0028] Figure 4 The diagram illustrates Gaussian noise and drift noise.

[0029] Figure 5 The diagram illustrates the VHI (Virtual Health Index) generator.

[0030] Figure 6 An example of the VHI boundary between the four degradation levels is shown.

[0031] Figure 7 A flowchart illustrating a method for estimating the lifespan index of a machine performing a given manufacturing process is shown according to an embodiment. Detailed Implementation

[0032] The following detailed description is taken with reference to the accompanying drawings, which illustrate by way of illustration specific details and aspects of the present disclosure in which the invention can be practiced. Other aspects may be utilized, and structural, logical, and electrical changes may be made, without departing from the scope of the invention. The various aspects of the present disclosure are not necessarily mutually exclusive, as some aspects of the present disclosure may be combined with one or more other aspects of the present disclosure to form new aspects.

[0033] The following sections will describe various examples in more detail.

[0034] Figure 1 A winding machine 100 is shown as an example of a manufacturing machine.

[0035] The winding machine 100 performs the winding process. The winding machine uses copper wire from a wire source 101, which is supplied to the winding unit 103 via a wire tensioner 102. The winding unit 103 winds the wire around the stator 104, which is transported by a conveyor platform 105. The winding machine typically has a second stage, which is a "connection" stage. For a given stator coil, once the copper wire has been wound around the stator (which has chambers that accommodate multiple stator coils), the conveyor platform 105 performs a rotational movement of the stator. This movement positions the next free stator coil slot under the needle rollers to begin the next winding stage of the stator.

[0036] For manufacturing plants to operate efficiently, maintenance of machinery such as the winding machine described above is essential. Predictive maintenance (PdM) is highly desirable because it allows for the early detection of potential equipment failures, reducing the likelihood of unexpected breakdowns. By using data and advanced analytics to predict when maintenance should be performed, manufacturers can schedule repairs at convenient times, minimizing downtime and avoiding costly interruptions. This proactive approach not only extends machine lifespan but also enhances safety, efficiency, and overall productivity. Additionally, predictive maintenance can lead to significant cost savings by preventing major repairs and optimizing resource usage.

[0037] However, many manufacturing plants and processes require a wide variety of machine types, all of which may have individual maintenance requirements. Therefore, a PdM framework that can be flexibly adapted to different machine types and sensor data is desirable.

[0038] In light of the above, a scalable method for PdM (specifically, RUL estimators) is provided, which can be applied to various machines and (e.g., sensors) data structures with minimal customization. PdM involves estimating the RUL of a machine, i.e., estimating the length of time the machine will operate before requiring repair or replacement. In other words, estimating RUL can be understood as predicting when the machine will require repair or replacement. The quality and availability of historical failure data can be used to improve such predictions, but it is not necessary to obtain reliable predictions within the provided framework.

[0039] According to various embodiments, the provided PdM method uses a machine learning PdM model with an architecture that processes time-series data collected from sensors and feeds this time-series data into various neural network layers.

[0040] Figure 2 The illustration shows a training pipeline for the PdM model, or in other words, the RUL (Remaining Useful Life) estimator, according to an embodiment.

[0041] To train the PdM architecture for a machine performing a given manufacturing process (e.g., winding), a time series set of sensor data collected from one or more machines operating the manufacturing process is input as historical data 201 (i.e., each time series is a time series of sensor data from the machine operating the manufacturing process; it should be noted that this may be the same machine that should predict the RUL).

[0042] For time series without observable historical degradation data, noise can be injected in step 202 to add the historical data time series with degradation. The historical data can be preprocessed in step 203 (e.g., normalized sensor values, etc.).

[0043] Then, patterns are identified and anomalies are detected within the time series of historical data. For example, anomaly detection is performed based on the reconstruction error of a temporal convolutional network autoencoder 204 (when the sensor data is multivariate) and the reconstruction error of a long short-term memory autoencoder 205 (when the sensor data is univariate). Of course, these autoencoders are also trained from scratch, for example, by using sensor data immediately after major maintenance, assuming a state of optimal health. In this way, the autoencoders can learn what is "normal." During the use of these autoencoders for anomaly detection, the deviation is recorded as reconstruction error, thus indicating anomalies.

[0044] This is followed by, in step 206, creating health labels for all timestamps (i.e., time indices) of each time series in the historical data. The creation of health labels follows the Virtual Health Index (VHI) concept and is performed by a VHI generator, allowing for the quantification of degradation in the machines. Once health labels have been generated for the time series of historical data (i.e., corresponding time series for which health labels have been generated for each time series of sensor data), the next processing step is to define thresholds for various degradation levels in step 207, including Remaining Usable Life (RUL) boundaries (i.e., thresholds). If historical failure data is not available (i.e., historical data does not include time series of machines that actually failed during the collection of historical data), an unsupervised clustering algorithm 208 (e.g., a Gaussian mixture model) is used to extract the degradation boundaries.

[0045] In 209, the VHI prediction model (also referred to as the "prediction model" in this paper) is trained to extend the sequence of health labels into the future based on time series of health labels generated for time series of historical data. This training can be performed by presenting a VHI prediction model with a portion of the time series of health labels generated for time series of historical data and training it to complete the time series of health labels.

[0046] Figure 3 The diagram illustrates the following: Figure 2 The use of trained PdM models (also known as inference pipelines).

[0047] For a machine to estimate RUL (i.e., to perform PdM), the time series of sensor data up to the current operating time point 301 is converted into a health label sequence through data preprocessing 302, anomaly detection (i.e., reconstruction error determination) 303, and health label generation 304 (corresponding to 203, utilizing the anomaly detection of the autoencoder 204 or 205 used in the training and application of the VHI generator in 206). Then, in 305, this health label sequence is extended into the future using a trained VHI prediction model.

[0048] Whenever a sustained sequence of health labels reaches the RUL threshold, the time difference up to that point in 306 is presented as the final output PdM architecture. This is the estimated RUL value. A key feature of this framework is its adaptability: the PdM architecture can operate with minimal human involvement and uses both labeled and unlabeled data. According to various embodiments, this framework combines Long Short-Term Memory (LSTM) autoencoders (AEs), Temporal Convolutional Networks (TCNs) AEs, and hybrid models to develop a reliable framework for VHI generation and RUL estimation for machines.

[0049] The following provides details about the PdM model (or framework), particularly regarding its training (e.g.) Figure 2 (as illustrated in the figure) and the use of this PdM model in RUL estimation (e.g.) Figure 3 (As shown in the illustration).

[0050] As per reference Figure 1 In the case of the described winding machine, sensor data includes, for example, the change in acceleration of the winding needle as it moves up and down. However, sensor data may also include force measurements (e.g., the force applied to the wire).

[0051] Noise injection can be performed due to the lack of historical failures in the historical data. In addition to the original historical data, a time series with injected noise can be included in the historical data. For example, a PdM model uses health data to train an anomaly detection component including an autoencoder; noise injection allows for the training of a VHI generator, enabling PdM operation even under sparse data conditions (e.g., new machines or very limited data). Noise injection occurs relative to the time series, simulating chronological degradation. Therefore, in such a case, noise injection can be considered as a sub-segment from the latest segment of health data, appended to an earlier segment of health data.

[0052] Examples of noise injected into the training time series (at least for some that do not show degradation) include drift noise and Gaussian noise.

[0053] Figure 4 The illustration shows Gaussian noise 401 and drift noise 402 (e.g., as applied to a time series of acceleration measurements).

[0054] Gaussian noise, combined with random noise following a normal distribution, causes random but primarily small fluctuations. Drift noise adds gradual changes that become larger over time, forming a trend. Therefore, it allows for slow, continuous changes in data values, effectively simulating gradual wear or degradation, which better reflects machine degradation and is thus usable. For example, in the case of a winding machine, acceleration values ​​typically increase with increasing machine degradation.

[0055] As described above, for anomaly detection, the framework includes two variants: an autoencoder 204 with a TCN layer or an autoencoder with an LSTM layer. This allows for the use of an LSTM-based autoencoder in the univariate case and a TCN-based autoencoder in the multivariate case due to better spatial feature extraction.

[0056] The reason for using the LSTM-based AE 205 is its superior performance in capturing temporal information, thanks to its adjustable memory, which has gating as its control mechanism. In the univariate case, sensor data consists of a single feature (e.g., acceleration) for each time stamp. These time stamps are grouped sequentially while the temporal order is preserved. The reason for using sequence with LSTM is simply the constraint of their internal memory. Although LSTMs are known for their ability to allow long-range temporal relationships by utilizing their gating and memory mechanisms, the length of a time stamp, the length across which information is continued, is not unlimited. Although there is no clear boundary, after a certain number of time stamps, the learning capacity in the internal memory becomes inefficient. This inefficiency usually does not lead to errors during code execution but tends to manifest itself in the evaluation of the training process. Therefore, to maximize both the length of internal memory seen each time and the learning efficiency, the sequence length can be set to a value such as 200. Another aspect to consider is the computational efficiency of the LSTM layers. Unlike other architectures that allow parallel processing of different time stamps (such as CNNs), LSTMs process time stamps sequentially. Unfortunately, this doesn't scale well with increasing data size. To offset this problem, batch parallelism can be used. By using an additional parameter called "batch size," this framework allows control over the order in which numbers are processed in parallel. While sequential behavior remains within individual sequences, sets of sequences are processed in parallel, forming batches. The weights of the AE network are then updated after each batch. Therefore, properly controlling the batch size is of great importance. A large batch size means processing more data at once, resulting in fewer weight updates on the network, but in return, faster training time will be achieved. If a smaller batch size is used, the desired behavior is reversed. Regarding the actual layer setup of the LSTM-AE, the layers in both the encoder and decoder parts can be mirrored, in other words, symmetrical, given that the purpose of the AE here is to highlight deviations through accurate reconstruction, which is essential for anomaly detection. Given its ability to capture both lower and higher temporal capabilities, stacked LSTM layers can be used for both the encoder and decoder parts. While this setup may not be strictly required for less complex use cases, the trade-off between the additional computational overhead and the improved generality of the proposed architecture may be worthwhile given scalability. Furthermore, according to various embodiments, these layers are designed to initially increase the dimensions to a user-defined value. This increases complexity, which is crucial for the basic network to extract the necessary patterns. A second layer then halves the dimensions to slowly prepare the data for the impending compression in the bottleneck layer at the end of the encoder section. The decoder then mirrors these steps to obtain the initial input dimensions at the end of the decoder.To guide the training process, for example, MSE (mean squared error) is chosen as a metric for reconstruction loss. Therefore, the reconstruction loss between the original order and the reconstructed order can be tracked, with the goal of minimizing them (for healthy time series as historical data). The choice of MSE here is reasonable because the penalty for stronger biases provides an additional value, allowing them to be included without additional attention to smaller errors that may be mixed with natural noise in the data. To further enhance the model's functionality, dropout layers might be included to help combat overfitting and improve generalization. Early stopping can be introduced to aid in epoch control.

[0057] According to various embodiments, for multivariate sensor data (e.g., including acceleration and force values), the TCN-AE 204 is used as internal memory for LSTM layers, where the potential for maximizing spatial feature extraction is not maximized due to the primary focus on the temporal aspect. The TCN-AE 204 includes one-dimensional convolutional layers and then treats features as individual channels, allowing the TCN-AE 204 to learn patterns simultaneously across multiple features. By using convolutional operations, both intra-feature and cross-feature temporal patterns can be learned, thereby capturing spatial dependencies that change over time. Because the TCN architecture does not process data sequentially like LSTMs, its runtime is generally faster, despite the increased data size due to the multivariate data format. Two important aspects of the TCN-AE 204 are the expansion property of the convolutional layers and the causal relationship of the processing order. Expansion is a technique that expands the kernel by inserting spaces between consecutive elements. This allows the TCN-AE 204 to expand its receptive field exponentially with each additional layer without causing a dramatic increase in the number of parameters.

[0058] According to various embodiments, the expansion starts from one layer and doubles with each subsequent layer. In other words, the expansion allows the TCN-AE 204 to "look further back in time" from a given point. Each output point in the sequence depends on a broader historical range of the input. This makes a larger sequence, compared to LSTM-AE 205, necessary to avoid underutilizing the expanded layers. On the other hand, causality refers to ensuring temporal order while processing data by guaranteeing no data leakage from the future. This can be achieved through sequence-based batch processing, similar to the processing in LSTM-AE 205. Sequence-based batch processing ensures that each sequence (i.e., time series) is confined to its own historical context. Zero padding can be used to ensure that the output sequence (i.e., the reconstructed time series) maintains the same length as the input sequence of the autoencoder. This becomes particularly important in the case of expanded convolutional layer settings and provides stability in terms of dimensionality rather than a direct impact on the causality of batch parallelism.

[0059] According to various embodiments, the encoding layer of TCN-AE 204 initially has three convolutional layers, each paired with a ReLU activation function. Within each convolutional layer, the input and output channels, kernel size, dilation, and padding are set. The input channels are equal to the input size of the first layer. For subsequent layers, this value is set to the desired number of filters, which is user-controlled and remains fixed across subsequent layers. Ideally, this value is set higher than the initial count of the sensors to allow for better feature extraction in higher dimensions. The output size remains equal to the number of filters, thus maintaining this dimension throughout the encoding layer. The final component of the encoding layer is a 1×1 convolutional operation. While it does not change the dimension, it serves as a thinning layer by applying a linear transformation across channels. It effectively functions as a "channel mixer" layer, which improves the network's ability to capture more complex interactions across multivariate data features. Similar to LSTM-AE 205, in TCN-AE 204, the decoder is symmetric to the encoder.

[0060] The training function is kept similar to that in the univariate case: MSE is chosen as the metric to calculate the loss (based on health time series) for training TCN-AE204.

[0061] Dilated convolutions and parallel processing make TCN-AE 204 more efficient than LSTM 205 because TCN-AE 204 avoids sequential dependencies, thus enabling faster training and inference. The dilated convolutional layout allows the model to capture both short-term and long-term dependencies without increasing computational load. On the other hand, the high parameter count of multiple convolutional layers can lead to overfitting on smaller datasets. If more layers are desired in the architecture, adding skip connections can help with gradient flow.

[0062] Figure 5 The illustration shows the VHI generator 500 used to generate health labels 206 and 303.

[0063] The VHI generator 500 uses the reconstruction error (also referred to as "raw anomaly") from the time-series input 501 (of the training set during training or the machine whose RUL should be estimated) to generate the reconstruction error.

[0064] During training, the VHI generator 500 learns machine-specific degradation patterns by training two (small) neural networks 503 and 504, respectively, called "k_anomaly_net" and "k_window_net". Training is guided by a (custom) loss function 508.

[0065] According to an embodiment, k_anomaly_net 503 is a small neural network consisting of two fully connected layers that directly uses the reconstruction error 501. The second network (k_window_net 504) relies on the same input 501, but instead of directly using the reconstruction error, it creates non-overlapping windows 502 within the input 501 (i.e., within the sequence of reconstruction errors corresponding to the timestamps of the respective time series), and k_window_net 504 averages the reconstruction errors within each of these windows to represent trends in a specific region. The window length corresponds to the batch size used in the autoencoder to allow the model to be fully autonomous and avoid hard-coding across different scenarios. Thus, even if outliers exist in the original anomaly 501 for a given data point, the window anomaly analysis performed by k_window_net 504 examines neighboring regions and enriches the context of the corresponding window.

[0066] Unprocessed raw outliers are input to a first neural network 503, while processed windowed outliers are used as input to a second neural network 504. From a deep learning perspective, the two neural networks 503 and 504 can have the same structure. According to various embodiments, each neural network has two fully connected (FC) layers with a ReLU activation function between these FC layers. Since the input is a scalar value of data points, the neural network increases the dimension from 1 to 16 in the first layer using a non-linear activation function, and then reduces the dimension back to 1 in the second FC layer. Forward pass is simply passing the correct input through the corresponding network. According to one embodiment, the final sigmoid is used at the output of the second FC layer. It stabilizes the created VHI, which requires stability, to realistically represent the slow but gradual wear and tear in a machine.

[0067] Therefore, the result of the processing performed by the two neural networks 503 and 504 can be written as: .

[0068] These results are referred to as “K values” in the following text, and are then processed as follows to generate VHI values.

[0069] In the 505 error correction, the weighted sum is determined as follows: the K values ​​are multiplied by their corresponding input scores, and then the two multiplication outputs are summed together. The result of this operation is denoted by "weighted_sum" because it effectively becomes a weighted sum of the original outlier and the windowed outlier values, as follows: Where x is the reconstruction error of the current timestamp, and y is the average result within the current window.

[0070] In 506, the weighted sum is added cumulatively. The result of this is called "cumulative_sum".

[0071] In 507, the VHI is calculated by subtracting the cumulative sum from the initial starting point of 1. This is performed for each time step (i.e., the time index of the corresponding time series), resulting in the generation of the VHI for each time step.

[0072] Therefore, with each pass (i.e., for each time step), the monotonic behavior of the VHI generator is guaranteed due to the cumulative sum. Using a sigmoid function and a custom loss function 508 helps to further stabilize the generated values.

[0073] According to various embodiments, the loss function 508 used to train the VHI generator has three components: • Initialize penalty: For the first VHI value, strictly penalize deviations from the scalar value 1 to ensure a healthy start.

[0074] • Stability penalty: Take the difference between adjacent VHI pairs and penalize their sum to ensure smoothness and high stability, both of which are important.

[0075] • Zero proximity penalty: In cases where there is a tendency during training for values ​​to prematurely reach zero, this component penalty becomes a value close to zero to ensure that these values ​​are genuinely low due to significant machine degradation.

[0076] The three components of the loss function 508 (used for training) can be written as follows: .

[0077] Table 1 provides an example of the time series processed by both the autoencoder and the VHI generator, where scaled acceleration values, original anomalies, window anomalies, and learned VHI values ​​can be seen. It should be noted that the VHI values ​​successfully start from an initial value of 1 and decrease slowly and monotonically.

[0078] Table 1.

[0079] According to various embodiments, when the PdM framework is first used on a new machine, the calibration mechanism 509 can be used for a certain period of time (e.g., two weeks or one month). This calibration mechanism allows the PdM model (e.g., neural networks 503, 504) to become accustomed to these new operating conditions under which it will operate without interruption. In this way, regardless of how high or low the sensor sampling rate is, and regardless of how long it takes for parts to be replaced on the previous machine, the VHI generator 500 (particularly including networks 503, 504) can accumulate experience and thus adapt the VHI values ​​it generates to be more accurate. This optimization step of the framework is possible by simply introducing an additional variable called “scaler_value”, which is placed inside the forward pass when calculating the weighted sum. Its job is to scale the value of the weighted sum. Initialized to 1 so that it has no effect, it accelerates or decelerates the rate at which health degradation is reflected in the virtual health index (also simply referred to herein as the “health index”).

[0080] The trained VHI generator is used in the evaluation pattern (i.e. for inference) to generate health labels for new data points in real-time production, allowing the degradation of VHI values ​​to be represented in an accurate manner.

[0081] As referenced above Figure 2 As explained in 207, various degradation levels are determined (i.e., thresholds between multiple degradation levels based on the VHI threshold), particularly the RUL threshold.

[0082] There are two scenarios: historical data includes (sufficient) historical data, or historical data does not include (sufficient) historical data, where the (sufficient) historical data includes the failure event (or the entry point into a specific degradation range) and the data available prior to that failure event. In the first case, the VHI values ​​output by the VHI generator for historical data including the failure (or the entry point into a specific degradation range) (i.e., the time series of historical data) can be used to define VHI thresholds for various degradation levels.

[0083] In the second case, an unsupervised ML algorithm (corresponding to the unsupervised clustering algorithm 208) can be used to define the threshold, i.e., to find the degenerate grouping in the generated VHI values. In this way, three to five groups (i.e., degenerate ranges) can be obtained, representing various levels of degradation. An example case with four groups can be labeled: healthy, slightly degraded, moderately degraded, and critically degraded. In this example, the RUL can be set as the boundary immediately preceding the critically degraded stage (i.e., the RUL threshold is the boundary between moderate and critically degraded). Although less reliable compared to failures occurring in real life (i.e., using historical data as in the first case), this approach allows the PdM framework to operate and function even in low-intelligence scenarios, generating potentially valuable insights that can be further improved later.

[0084] Examples of unsupervised clustering algorithms 208 are the K-means clustering algorithm, Gaussian mixture model, and DBSCAN (density-based spatial clustering for noisy applications) clustering algorithm. GMM assigns probabilities to clusters, i.e., the probability that a given sample belongs to a given cluster.

[0085] Figure 6 An example of VHI boundaries 601-603 between the four degradation levels found using GMM is shown. At each boundary 601-603, the probability of belonging to any cluster (degradation level) is equal. The lowest boundary 603 identifies the RUL threshold within the VHI value. This approximately corresponds to the region where moderate degradation becomes critical degradation.

[0086] exist Figure 6 In the diagram, the VHI value increases along the y-axis, and the time step is indicated along the x-axis.

[0087] An example of a VHI prediction model trained in 209 is Prophet, an open-source time series forecasting tool originally developed by Facebook (Meta). It uses additive regression, which decomposes data into sub-components such as trend, seasonality, holidays, and an error term. It offers additional flexibility by allowing users to add custom features while remaining robust to noise and outliers in the data. Since the PdM framework requires that future VHI values ​​consist only of the VHI values ​​themselves, which are essentially single columns or features, Prophet becomes a strong candidate for this purpose. According to one embodiment, when the framework is applied to manufacturing machines, the seasonality and holiday components are omitted, further simplifying the VHI prediction model. Therefore, the remaining components are simply the trend in the data and the error term adjacent to the trend.

[0088] The input to the predictive model (e.g., in the case of Prophet) is a data frame with two columns: a timestamp column and a column for numerical target features; in this case, the data frame is the VHI. It should be noted that if the historical data is healthy, Prophet cannot know how the pattern will look when degradation begins. Since it is trained only on healthy data, it effectively only perpetuates healthy trends into the future. This trap is avoided by showing a declining VHI trend, either from real-world failures or from injected noise, so that it sees in a machine-specific way how it should fit its future predictions. Once this information is fed in (i.e., the VHI predictive model is trained, in this case, Prophet), the VHI values ​​are determined using a VHI generator based on current sensor data under the evaluation pattern (i.e., for inference), and the VHI values ​​are perpetuated into the future using the trained predictive model. To keep the VHI prediction model fed with more and more up-to-date information, incoming sensor data can be continuously used to generate additional VHI values. These VHI values ​​are then appended to the VHI values ​​generated for the machine under consideration (i.e., for which a RUL should be determined), thus providing the VHI prediction model with a broader context for making predictions. This ensures that the predicted VHI values ​​become more accurate as the sensor data matures.

[0089] If the machine under consideration is healthy and reaching the RUL threshold within a satisfactory timeframe is not foreseeable, RUL predictions can be set to automatically terminate based on a time cap (i.e., a future time index). This cap can be selected by the user as needed, as different spare parts may require different timeframes for data collection in a production setup. While ensuring no RUL is reached within a few days might be sufficient for one user, another user might benefit from generating more values ​​if forecasts for two weeks or longer are available. Therefore, the upper limit of the forecast time range is left to the user for flexible adjustment.

[0090] If the predicted VHI value does not meet the RUL threshold within the desired timeframe (i.e., before the time limit), this is the final output of the PdM framework in this case. However, if the predicted VHI value does meet the RUL threshold, the RUL estimate is calculated simply for the number of time points before the predicted VHI value reaches the RUL threshold. In effect, this is a subtraction of two values. Maintenance actions, such as machine or component replacement, machine outages, activation of safe mode, etc., can then be triggered based on the estimated RUL.

[0091] In summary, according to various embodiments, the following are provided: Figure 7 The method is illustrated in the figure.

[0092] Figure 7 A flowchart 700, illustrating a method for estimating the lifespan index of a machine performing a given manufacturing process, is shown according to an embodiment. Estimating the lifespan index can be understood as estimating the time until an event related to the machine's lifespan occurs, such as until the machine fails at a certain level or its operational quality deteriorates.

[0093] In 701, historical data including at least one historical data time series is received, wherein the at least one historical data time series is a time series of sensor data collected during the execution of the manufacturing process.

[0094] In 702, for at least one historical data time series of historical data, a training time series of a health index for a specified machine's health is generated by applying an anomaly detector to the historical data time series and using a health index generator to determine the health index of the training time series based on the anomaly detection results of the anomaly detector.

[0095] In 703, a range of health indices representing different levels of degradation and one or more thresholds between a pair or more pairs of degradation levels are generated, including a threshold representing the end of lifespan.

[0096] In 704, the generated health index training time series is used as training data, and the health index prediction model is trained to extend the health index time series into the future.

[0097] In 705, for machines performing manufacturing processes, machine life indicators are estimated based on time series of sensor data collected during machine operation, using the following: • By applying an anomaly detector to the time series of sensor data collected during machine operation and using a health index generator, a machine health index time series is generated based on the time series of sensor data collected during machine operation. • Use trained health index prediction models to extend the machine's health index time series into the future, and • Determine the estimated lifespan index of the machine from the point in time when the machine's ongoing health index time series reaches the threshold indicating the end of its service life.

[0098] It should be noted that this order is not necessarily as follows. Figure 7 As indicated in the document. In particular, the predictive model can be executed before or during the determination of the range and thresholds.

[0099] For example, a method is provided for estimating life indicators (such as remaining useful life) of a machine performing a given manufacturing process, comprising receiving time-series data collected during the machine's operation from sensors on the machine. The received data is then fed to an anomaly detection engine (ADE) to understand anomalies in the machine's behavior. Here, patterns of normal machine behavior are identified by the engine, and any deviations indicate that the machine is undergoing wear and tear during its operation. The output of the ADE is a machine health index indicating a trend of degradation. The method then includes determining a range of health indices representing different levels of degradation, including a final threshold indicating the machine requires maintenance at the end of its useful life, and training a health index prediction model using existing health indices to extend the health index time series into the future, thus knowing when the machine will reach its remaining useful life.

[0100] It should be noted that historical data can also be regarded as training data for PdM frameworks (i.e., anomaly detectors, VHI generators, etc.).

[0101] Figure 7 The method can be used to calculate control signals for controlling manufacturing machines.

[0102] Sensor data can include acceleration, pressure, force, etc., but it also includes image data (i.e., digital images) from various visual sensors (cameras) (such as video, radar, LiDAR, ultrasound, thermal imaging, motion, sonar, etc.).

[0103] Figure 7 The method can be performed by one or more data processing devices (e.g., a computer or controller) having one or more data processing units. The term "data processing unit" can be understood to mean any type of entity that implements the processing of data or signals. For example, data or signals can be addressed according to at least one (i.e., one or more) specific functions performed by the data processing unit. The data processing unit can include or be formed from analog circuits, digital circuits, logic circuits, microprocessors, microcontrollers, central processing units (CPUs), graphics processing units (GPUs), digital signal processors (DSPs), field-programmable gate arrays (FPGAs), or any combination thereof. Any other means for implementing the corresponding functions described in more detail herein can also be understood to include data processing units or logic circuits. One or more of the method steps described in more detail herein can be performed (e.g., implemented) by the data processing unit through one or more specific functions performed by the data processing unit.

[0104] Therefore, according to one embodiment, the method is implemented by a computer.

Claims

1. A method for estimating the life index of a machine (100) performing a given manufacturing process, comprising: Receive (701) historical data including at least one historical data time series, the historical data time series being a time series of sensor data collected during the execution of the manufacturing process; An anomaly detector (702) is applied to at least one historical data time series of the historical data, and a health index generator (500) is used to generate a training time series of a health index specifying the health of the corresponding machine (100) based on the anomaly detection results of the anomaly detector. Determine (703) a range of health indices representing different levels of degradation and one or more thresholds between a pair or more pairs of degradation levels, the one or more thresholds including a threshold representing the end of life (603). The generated health index training time series is used as training data to train the (704) health index prediction model to extend the health index time series into the future; For the machine (100) performing the manufacturing process, the lifespan index of the machine (100) is estimated (705) based on the time series of sensor data collected during the operation of the machine (100) using the following: The anomaly detector is applied to the time series of sensor data collected during the operation of the machine (100) and the health index time series of the machine (100) is generated based on the time series of sensor data collected during the operation of the machine (100) by using the health index generator (500).    Using a trained health index prediction model, the time series of the health index of the machine (100) is extended into the future; as well as The estimated lifespan index of the machine (100) is determined from the point in time when the continuous health index time series of the machine (100) reaches the threshold (603) indicating the end of its lifespan.

2. The method of claim 1, further comprising determining, by applying a machine learning cluster model, a range of health indices representing different levels of degradation based on the generated health index time series.

3. The method according to claim 1 or 2, comprising generating a historical data time series of the historical data by injecting noise (401, 402) representing degradation into a time series of sensor data collected during the execution of the manufacturing process.

4. The method according to any one of claims 1 to 3, wherein the health index generator (500) generates the health index by changing an initial health index representing a healthy machine over time based on the anomaly detection results.

5. The method according to claim 4, wherein the health index generator (500) includes a first neural network (503) that processes anomaly detection results of anomaly detection of the current time index and a second neural network (504) that processes anomaly detection results of anomaly detection averaged over a predetermined time index window, combines the outputs of the first neural network (503) and the second neural network (504), and determines the health index value of the current time index by changing the health index value from the prior time index according to the result of the combination.

6. The method of claim 5, comprising training the health index generator (500) by training the first neural network (503) and the second neural network (504) using a loss function, the loss function comprising at least one of a penalty for the deviation between a subsequent health index and a health index that closely represents the maximum degradation.

7. The method according to any one of claims 1 to 6, wherein the anomaly detector comprises an autoencoder (204, 205), and the result of the anomaly detector includes the reconstruction error of the autoencoder (204, 205).

8. The method of claim 7, wherein the autoencoder is a temporal convolutional network autoencoder or a long short-term memory autoencoder.

9. A method for controlling a machine (100), comprising estimating a life index of the machine (100) according to any one of claims 1 to 7, and applying maintenance actions to the machine (100) based on the estimated life index.

10. A data processing system configured to perform the method according to any one of claims 1 to 9.

11. A computer program comprising instructions that, when executed by a computer, cause the computer to perform the method according to any one of claims 1 to 9.

12. A computer-readable medium comprising instructions that, when executed by a computer, cause the computer to perform the method according to any one of claims 1 to 9.

Images