Intelligent early warning method and system for sintering furnace of powder metal metallurgical part

Abstract

This invention relates to the field of data processing, specifically to an intelligent early warning method and system for sintering furnaces used in powder metallurgy. The method includes: collecting historical normal operation data and dividing it into multiple windows; extracting multi-dimensional features for each window, including a structural feature vector based on the singular values ​​of the autocorrelation matrix, a coupling feature vector based on the cross-correlation matrix, an inertial-coupling dominant feature, and an inertial-coupling consistency vector; using the structural feature vector as the first sub-vector and the remaining features as the second sub-vector, employing a cosine kernel and a radial basis function kernel respectively, and determining the weights of the two kernels to construct a combined kernel function to train a single-class support vector machine; collecting data in real time and extracting identical features, and substituting them into a decision function to determine whether an early warning is triggered. This invention, through multi-dimensional feature extraction and adaptive combined kernel functions, can effectively distinguish between normal fluctuations and abnormal precursors, improving the accuracy of early warnings for sintering furnace anomalies.

Description

Technical Field

[0001] This invention relates to the field of data processing. More specifically, this invention relates to an intelligent early warning method and system for sintering furnaces of powder metallurgy parts. Background Technology

[0002] The sintering process of powder metallurgy parts is usually carried out in a sintering furnace. The compact is heated to a temperature close to the melting point of the main components under a protective atmosphere, and the particles are diffused and bonded together by heat preservation, so as to obtain a dense metallurgy part.

[0003] The stability of parameters such as temperature, furnace pressure, and protective gas flow rate directly determines product quality. Currently, sintering furnaces widely adopt a pre-alarm method based on fixed thresholds, which means that a static upper and lower limit is preset for each process variable, and an alarm is triggered when the real-time measured value exceeds the range.

[0004] However, the sintering process is a continuous thermophysical dynamic process. Even under normal operating conditions, a single parameter can drift slowly or oscillate slightly due to different sintering stages and furnace condition fluctuations. Fixed threshold methods struggle to distinguish between normal dynamic fluctuations and true precursors of anomalies, often leading to delayed warnings or frequent false alarms, and failing to accurately identify early abnormal states of the sintering furnace. Summary of the Invention

[0005] To address the technical problem that the fixed threshold method in the prior art cannot accurately identify early abnormal states of sintering furnaces, the present invention provides solutions in the following aspects.

[0006] In the first aspect, a method for intelligent early warning of sintering furnaces for powder metallurgy parts includes: Collect historical normal operation data of the sintering furnace and preprocess it to obtain a set of historical windows consisting of a sliding window of normal operating conditions; Multidimensional features of each window are extracted from the set of historical windows. The multidimensional features include: structural feature vectors extracted based on the singular values ​​of the autocorrelation matrix of each parameter, coupling feature vectors extracted based on the cross-correlation matrix between each parameter, and inertial-coupling dominant features and inertial-coupling consistency vectors extracted based on joint analysis of autocorrelation and cross-correlation. The extracted multidimensional features are used to form a training sample set. The multidimensional features of each training sample are divided into a first sub-vector and a second sub-vector. The first sub-vector is the structural feature vector, and the second sub-vector is a vector composed of features other than the structural feature vector. The kernel function corresponding to the first sub-vector is a cosine kernel, and the kernel function corresponding to the second sub-vector is a radial basis function kernel. The weights of the cosine kernel and the radial basis function kernel are determined according to the mean of the inertial-coupling dominant features of all training samples in the training sample set. The combined kernel function obtained by weighted summation is used to train a single-class support vector machine to obtain the decision function. The system collects real-time operating data of the sintering furnace, extracts multi-dimensional features from the real-time window, and substitutes them into the decision function of a pre-trained single-class support vector machine for discrimination. When the decision function value indicates an anomaly, an early warning is triggered.

[0007] Optionally, the historical normal operation data includes at least time-series data of three parameters: temperature, pressure, and flow rate. The preprocessing includes using a sliding window to divide the long time series into overlapping short time windows and normalizing the data within each window.

[0008] Optionally, the structural feature vector is extracted in the following way: Select any window in the historical window set as the target window. For the target window, set the maximum lag order, calculate the autocorrelation coefficient of each parameter at each lag order and arrange them into an autocorrelation matrix. Perform singular value decomposition on the autocorrelation matrix to obtain multiple singular values. Concatenate the multiple singular values ​​in descending order to form a structural feature vector.

[0009] Optionally, the coupling feature vector is extracted in the following way: For the target window, the same lag order is used. For each lag order, the cross-correlation function between any two parameters is calculated and arranged into a cross-correlation matrix to obtain the cross-correlation matrix for each lag order. For each directed parameter pair set according to historical normal operation data, the maximum value of the absolute value sequence of its cross-correlation values ​​under all lag orders is extracted, and the maximum values ​​corresponding to all directed parameter pairs are arranged in a fixed order to form a coupling feature vector.

[0010] Optionally, the extraction method of the inertial-coupling dominant feature is as follows: For each lag, the mean of the absolute values ​​of the autocorrelation coefficients is calculated as the average inertial strength, and the mean of the absolute values ​​of all off-diagonal elements in the cross-correlation matrix is ​​calculated as the average coupling strength. Then, the inertial-coupling dominance ratio of the lag is obtained according to the relationship between the average inertial strength and the average coupling strength. The mean of the inertial-coupling dominance ratios of all lags is taken as the inertial-coupling dominance feature.

[0011] Optionally, the inertial-coupling consistency vector is extracted as follows: For each parameter, the full-scale inertia vector corresponding to the parameter is extracted from the autocorrelation matrix, and the full-scale coupling vector corresponding to the parameter is extracted from the cross-correlation matrix group. The cosine similarity between the full-scale inertia vector and the full-scale coupling vector is calculated, and the cosine similarity of all parameters is concatenated to form the inertia-coupling consistency vector.

[0012] Optionally, the weight of the cosine kernel is obtained by truncating the mean of the inertial-coupling dominant features of all training samples in the training sample set to a preset interval, and the weight of the radial basis kernel is equal to 1 minus the weight of the cosine kernel.

[0013] In a second aspect, an intelligent early warning system for a sintering furnace of powder metallurgy parts includes: a processor and a memory, wherein the memory stores computer program instructions, and when the computer program instructions are executed by the processor, the intelligent early warning method for a sintering furnace of powder metallurgy parts described in any one of the claims is implemented.

[0014] The present invention has the following beneficial effects: 1. This invention employs multi-dimensional feature extraction from normal sintering furnace operation data, including parameter inertia, inter-parameter coupling relationships, and overall dynamic bias, for discriminative processing. Based on the dominant inertia-coupling features of the sample data itself, it adaptively matches appropriate kernel functions and weights to different feature subspaces, thereby constructing a combined kernel function to train a single-class support vector machine. This approach enables the early warning model to sensitively capture the dynamic boundaries of normal operating conditions, effectively distinguishing normal parameter fluctuations from true anomalies, thus significantly reducing the false alarm and false negative rates common in fixed threshold methods, achieving accurate and timely early warning of early abnormal states in the sintering furnace.

[0015] 2. This invention extracts structural features through singular value decomposition of the autocorrelation matrix, obtains coupling features by combining cross-correlation matrix analysis of successive hysteresis, and further constructs the inertia-coupling dominance ratio and inertia-coupling consistency vector. It characterizes the system operation state from two complementary dimensions: overall bias and individual morphological similarity, which enhances the model's adaptability to the complex thermophysical dynamic processes of the sintering furnace and further improves the robustness and accuracy of early anomaly identification. Attached Figure Description

[0016] Figure 1 This is a flowchart of steps S1-S4 in an intelligent early warning method for sintering furnaces of powder metallurgy parts according to an embodiment of the present invention.

[0017] Figure 2 This is a schematic diagram illustrating the process of extracting multi-dimensional features from pre-processed historical normal operation data in an intelligent early warning method for sintering furnaces of powder metallurgy parts according to an embodiment of the present invention.

[0018] Figure 3 This is a structural block diagram of an intelligent early warning system for a sintering furnace of powder metallurgy parts according to an embodiment of the present invention. Detailed Implementation

[0019] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are some embodiments of the present invention, but not all embodiments.

[0020] Reference Figure 1 A method for intelligent early warning of sintering furnace for powder metallurgy parts includes steps S1-S4, as detailed below: S1: Collect historical normal operation data of the sintering furnace and perform preprocessing.

[0021] S2: Extract multidimensional features from preprocessed historical normal operation data.

[0022] S3: The extracted multidimensional features are used to form a training sample set and a multi-kernel single-class support vector machine is trained.

[0023] S4: Use a trained multi-core single-class support vector machine to determine the sintering furnace operating conditions online and trigger an early warning.

[0024] The above method will be described in detail below with reference to specific embodiments.

[0025] S1: Collect historical normal operation data of the sintering furnace and perform preprocessing.

[0026] During normal and stable operation, the sintering furnace exhibits certain temporal variations in parameters such as temperature, pressure, and protective gas flow rate. To establish a baseline model for normal operating conditions, it is first necessary to collect a sufficient amount of historical normal data and divide it into several time windows. The short-time series matrix of each window can both preserve local dynamic characteristics and adapt to the long-term non-stationarity of the sintering process. The purpose of the sliding window operation is to increase the number of samples while ensuring overlap between adjacent windows, thereby avoiding information gaps.

[0027] In an actual powder metallurgy production line, the sintering furnace is a mesh belt continuous furnace. First, a K-type thermocouple is installed in the middle of the furnace chamber to collect temperature data, a pressure transmitter is installed at the furnace exhaust pipe to collect furnace pressure data, and a thermal mass flow meter is installed on the main protective gas (ammonia decomposition gas) pipeline to collect instantaneous flow rate data.

[0028] The three sensors start synchronously at the same sampling frequency, such as 100Hz, and the temperature, pressure, and flow data at each sampling moment are completely aligned through a high-precision clock synchronization protocol such as IEEE1588. Data from more than 40 batches of normal sintering processes are continuously collected, with each batch lasting approximately 90 minutes from feeding to unloading.

[0029] Then, the length and step size of the sliding window are set. Considering the inertia of sintering furnace parameter changes, each window should contain enough sampling points to reflect short-term dynamics, while not being too long to avoid masking local fluctuations. In this embodiment of the invention, the sliding window length is set to 300 sampling points, corresponding to 3 seconds, and the step size is set to 100 sampling points, corresponding to 1 second. Starting from the beginning of each batch of data, the sliding window is continuously adjusted with the step size to divide the continuous long-term series into a series of overlapping short-term windows. The data in each window is organized into a matrix with 300 rows and 3 columns, including temperature, pressure, and flow rate. Each element in the matrix is ​​the original measurement value of the corresponding sensor at the corresponding time.

[0030] The raw data needs to be normalized to eliminate the influence of different units. In this embodiment of the invention, conventional Min-Max normalization is used. All normalized values ​​are within... The intervals preserve the relative changes while facilitating subsequent matrix operations. The above normalization is performed on each normal window, and the summaries yield the historical window set.

[0031] S2: Extract multidimensional features from preprocessed historical normal operation data.

[0032] The above-mentioned S1 yielded a set of historical windows composed of sliding windows under normal operating conditions. Each window in this set contains short-time synchronous sequences of three parameters inside the sintering furnace: temperature, pressure, and flow rate. However, these raw sequences cannot directly reveal two key types of information: first, the temporal inertia of each parameter itself, i.e., the influence of past values ​​on the current value; and second, the causal coupling relationships between different parameters, such as how furnace pressure fluctuations laggingly affect the protective gas flow rate. To simultaneously mine these two fundamentally different types of information, it is necessary to construct autocorrelation and cross-correlation matrices separately, and further extract multidimensional features using their shared hysteresis characteristics.

[0033] Reference Figure 2 The extraction process of the above-mentioned multidimensional features includes S20-S23.

[0034] S20: Construct an autocorrelation matrix based on preprocessed historical normal operation data and perform singular value decomposition to extract structural feature vectors.

[0035] The time inertia of a single parameter can be described by the autocorrelation coefficients at different time delays (lag orders), representing the number of steps between the current moment and a past moment. If all autocorrelation coefficients corresponding to all lag orders are directly used as features, the feature dimensionality becomes too high and redundant. The system needs to first construct an autocorrelation matrix based on the three types of parameter sequences within the target window, and then perform singular value decomposition to reduce its dimensionality and extract its main structural information.

[0036] First, select any window from the historical window set of S1 above as the target window, and obtain the temperature time series, pressure time series and flow time series within the target window.

[0037] Secondly, set the maximum hysteresis order, for example, take one-third of the target window length, i.e., 100, and let the hysteresis order take values ​​of 1, 2, 3, ..., 100 in sequence.

[0038] For each lag step, the autocorrelation coefficients for temperature, pressure, and flow rate are calculated separately. It should be noted that the calculation of the autocorrelation coefficient is common knowledge and will not be elaborated upon here.

[0039] Next, the calculated autocorrelation coefficients for temperature, pressure, and flow rate corresponding to all hysteresis orders are arranged into a 3-row × 100-column autocorrelation matrix.

[0040] Furthermore, performing singular value decomposition on the above autocorrelation matrix yields three non-negative singular values: .in, It is the largest singular value, representing the energy or variance contribution of the autocorrelation matrix in the principal direction, reflecting the common principal trend of the three-parameter time-series inertial structure; These are intermediate singular values, representing contributions in the secondary principal directions and reflecting secondary characteristics of the inertial structure; These are the smallest singular values, reflecting the remaining details or noise components. The relative magnitudes of the three singular values ​​together describe the overall shape of the autocorrelation matrix: if Much larger This indicates that the inertial patterns of the three parameters are highly consistent; if the three values ​​are relatively close, it indicates that the inertial characteristics of each parameter are relatively independent. These three singular values ​​are concatenated in descending order to form a structural feature vector: .

[0041] S21: Construct cross-correlation matrices step by step based on the same lag order and extract coupled feature vectors.

[0042] Causal coupling between parameters needs to be quantified through cross-correlation. Cross-correlation also depends on the lag order, that is, the correlation between the current value of one parameter and the value of another parameter at a past moment. Since coupling may have directionality, for example, pressure changes may affect flow, but flow changes do not necessarily affect pressure, it is necessary to distinguish between the two different directions, called directed parameter pairs. For three parameters, considering all possible directions, there are a total of 6 directed parameter pairs. Directly using the cross-correlation of all lag orders as the feature dimension would be too high, so the same lag order is used to construct cross-correlation matrices order by order, forming a matrix group, and then the maximum absolute cross-correlation coefficient of each directed parameter pair over all lag orders is extracted as the coupling feature.

[0043] Taking the target window as an example again, firstly, we use the same hysteresis orders 1, 2, 3, ..., 100. For each hysteresis order, we calculate the cross-correlation function between any two parameters, such as temperature and pressure. It should be noted that the calculation of the cross-correlation function is common knowledge and will not be elaborated on here.

[0044] For a fixed lag, the cross-correlation values ​​calculated from any two parameter combinations are arranged into a 3×3 cross-correlation matrix. This process is repeated to obtain a set of cross-correlation matrices for each lag. For the three parameters of temperature, pressure, and flow rate, consider six directed parameter pairs: temperature to pressure, pressure to temperature, temperature to flow rate, flow rate to temperature, pressure to flow rate, and flow rate to pressure.

[0045] For each directed parameter pair, extract the absolute value sequence of its cross-correlation values ​​under all lags, and take the maximum value of the sequence. Taking the maximum value can capture the most significant potential influence intensity in that direction.

[0046] Furthermore, the maximum values ​​corresponding to the above six directed parameter pairs are arranged in the fixed order given above to form a coupled feature vector: , Indicates temperature. Indicates pressure, Indicates flow rate. This represents the maximum value of the absolute sequence of cross-correlation values ​​of the directed parameter pair from temperature to pressure across all hysteresis orders.

[0047] S22: The joint eigenvector is obtained by using the constructed autocorrelation matrix and cross-correlation matrix.

[0048] The autocorrelation matrix in the above S21 operation describes the temporal inertia of a single parameter, and the cross-correlation matrix group describes the coupling relationship between parameters. However, the two alone cannot answer two deeper questions: First, within a window, is inertia or coupling dominant overall? Second, is the inertial shape of each parameter consistent with the shape affected by other parameters?

[0049] To address the issues in the two dimensions mentioned above, it is necessary to utilize the shared hysteresis alignment property of the autocorrelation and cross-correlation matrices to extract inertia-coupling dominant features and inertia-coupling consistency features. The former reflects the overall strength of inertia and coupling, while the latter reflects the similarity between the inertial shape and the shape affected by coupling for each parameter. These two features complement each other and together constitute joint features.

[0050] Taking the target window as an example, firstly, for each lag step, the mean of the absolute values ​​of the autocorrelation coefficients of temperature, pressure and flow rate in the autocorrelation matrix above is calculated as the average inertia strength of that lag step. The larger the average inertia strength, the stronger the overall inertia of the three parameters under that lag step.

[0051] Then, for each lag order, the mean of the absolute values ​​of all off-diagonal elements in the cross-correlation matrix is ​​calculated as the average coupling strength of that lag order. The larger the average coupling strength, the stronger the overall coupling between parameters under that lag order.

[0052] Next, the ratio of the average inertia strength to the average coupling strength of a single hysteresis order is calculated and its natural logarithm is taken to obtain the dominant inertia-coupling characteristic at that hysteresis order, which satisfies the following relationship: In the formula, Lag order The inertial-coupling dominance ratio is below. Lag order The average inertial strength, Lag order The average coupling strength, To prevent division by zero of extremely small constants, such as , Let be a logarithmic function with the natural constant e as the base.

[0053] when > hour, >0 indicates that in the lag stage Inertia dominates; when < hour, <0 indicates that coupling is dominant; when ≈ hour, ≈0 indicates that the two are in balance. Then, the inertial-coupling dominance ratio is calculated for all hysteresis orders, and the average value is taken to obtain the inertial-coupling dominance feature within the target window.

[0054] Furthermore, taking temperature as an example, the absolute value of the row vector corresponding to temperature is extracted from the autocorrelation matrix to obtain the full-scale inertia vector. This full-scale inertia vector describes the decay or oscillation pattern of temperature's own inertia with respect to the hysteresis order. In addition, the off-diagonal elements of the first row are taken from the cross-correlation matrix of any hysteresis order, namely the cross-correlation value of pressure to temperature and the cross-correlation value of flow rate to temperature. The mean of the absolute values ​​of these two cross-correlation values ​​is calculated, and after traversing all hysteresis orders, the full-scale coupling vector is obtained. This full-scale coupling vector describes the distribution pattern of the intensity of temperature being affected by other parameters with respect to the hysteresis order.

[0055] Furthermore, the cosine similarity between the full-scale inertial vector and the full-scale coupling vector is calculated. The larger the value, the more consistent the inertial change pattern of temperature is with the pattern of its influence by pressure and flow. Similarly, the cosine similarity between the full-scale inertial vector and the full-scale coupling vector of the pressure parameter, and the cosine similarity between the full-scale inertial vector and the full-scale coupling vector of the flow parameter are calculated. These three cosine similarities are then concatenated to form the inertial-coupling consistency vector within the target window.

[0056] The inertial-coupling consistency vector differs from the aforementioned inertial-coupling dominant feature in that the inertial-coupling dominant feature compares the magnitude of the overall strength, while the consistency feature compares the similarity of the morphology. The combination of the two can more comprehensively distinguish between normal and abnormal conditions. For example, under normal operating conditions, the autocorrelation morphology of temperature is usually highly consistent with the morphology affected by coupling, while abnormal operating conditions such as sensor drift may cause the autocorrelation morphology to change but the coupling morphology remains unchanged, thus significantly reducing the consistency.

[0057] In summary, the above operations yielded a 1-dimensional inertial-coupling dominant feature and a 3-dimensional inertial-coupling consistency feature. These two joint features supplement the system coordination information that cannot be obtained from a single matrix, reflecting the overall bias and individual morphological consistency, respectively.

[0058] S23: Concatenate all features and normalize them to generate a multidimensional feature vector.

[0059] The features obtained from S20-S22 above have different sources and may have different dimensions. It is necessary to unify the scale of each feature to form the final multidimensional features of a single window.

[0060] The extracted structural feature vector, coupling feature vector, inertial-coupling dominant feature, and inertial-coupling consistency vector are all subjected to Min-Max normalization. Then, the normalized structural feature vector, coupling feature vector, inertial-coupling dominant feature, and inertial-coupling consistency vector are concatenated end to end to form a multidimensional feature with a total dimension of 3 + 6 + 1 + 3 = 13.

[0061] At this point, each window in the historical window set has been transformed into a 13-dimensional multi-dimensional feature.

[0062] S3: The extracted multidimensional features are used to form a training sample set and a multi-kernel single-class support vector machine is trained.

[0063] The above S2 yields a training sample set composed of 13-dimensional multi-dimensional features from the normal operating condition window. These samples should theoretically cluster in the feature space. However, the structural feature vectors in the 13-dimensional features are non-negative and have bounded magnitudes, making them suitable for measuring directional similarity using a cosine kernel; while the other features besides the structural feature vectors are ordinary Euclidean vectors, suitable for measuring local density using a radial basis function kernel. Therefore, in this embodiment of the invention, a dual-kernel weighted combination strategy is adopted, and the weight coefficients are determined by the inertia-coupling bias of the training data itself.

[0064] First, the multidimensional features corresponding to each training sample, i.e. each window, are divided into a first sub-vector and a second sub-vector. The first sub-vector is the structural feature vector, and the second sub-vector is a vector composed of features other than the structural feature vector.

[0065] Next, the cosine kernel of the first sub-vector between any two training samples is calculated, i.e., the cosine similarity between the first sub-vectors corresponding to these two training samples is calculated. A larger value indicates a more consistent autocorrelation structure direction between the two training samples. Additionally, the radial basis function (RBF) kernel of the second sub-vector between these two training samples is calculated. This RBF kernel is equal to an exponential function calculated by multiplying the square of the negative Euclidean distance between the second sub-vectors of the two training samples by a preset parameter, which is selected through cross-validation, for example, a value of 0.5. This kernel function monotonically decreases to 0 as the Euclidean distance between the second sub-vectors of the two samples increases, and is sensitive to local density.

[0066] Considering that under normal operating conditions, the overall operation of the sintering furnace may exhibit varying degrees of inertia or coupling, affecting the importance of the two types of features in classification, it is necessary to determine the weights of the two kernel functions based on the average value of the inertia-coupling dominant features across all windows in the training sample set.

[0067] Extract the mean of the dominant inertial-coupling features from all training samples in the training sample set. To avoid extreme weights, truncate this mean to a specified interval. Then the weight of the cosine kernel is: ,in Let be the mean of the inertial-coupling dominant features of all training samples in the training sample set. The maximum value, If the minimum value is reached, then the weight of the radial basis kernel is the difference between 1 and the weight of the cosine kernel.

[0068] After obtaining the weights of the two kernel functions, a weighted sum is performed to obtain a composite kernel function. A single-class support vector machine is then trained based on this composite kernel function to characterize the boundaries of normal operating conditions. The specific training process is as follows: Constructing the combined kernel function: For any two training samples, the combined kernel function is equal to the sum of the cosine kernel and the radial basis kernel multiplied by their respective weights. This linear combination is still a positive definite kernel, equivalent to the inner product of the concatenated feature spaces.

[0069] Training is performed using a ν-single-class support vector machine, a method well-known in the art. Assuming the number of training samples is N, a quadratic programming problem is solved by introducing slack variables and Lagrange multipliers to obtain an optimal classification hyperplane. This hyperplane ensures that normal samples fall on one side as much as possible, while allowing a maximum of ν proportions of training samples to fall on the other side, which are considered outliers. In this embodiment, ν = 0.05, meaning a maximum tolerance of 5% outliers. After training, a set of support vectors and their corresponding weight coefficients are obtained.

[0070] Obtaining the decision function: For any real-time sample to be judged, i.e., the multi-dimensional features extracted and normalized from the real-time window according to the S2 operation online, the decision function value is the sum of the combined kernel function values ​​of the Lagrange multipliers of all support vectors and the real-time sample, minus a threshold. This threshold is automatically obtained through dual optimization during training, and its specific value is the weighted sum of the kernel functions corresponding to any support vector. In actual calculation, the average value of the corresponding values ​​of all support vectors can be taken. Its function is to ensure that the classification hyperplane separates normal samples from the origin to the greatest extent. When the decision function value is greater than or equal to 0, the system determines the sample as normal; when it is less than 0, it determines it as abnormal.

[0071] This step completes the construction of the combined kernel function and the training of the single-class support vector machine, and the model can be used for online discrimination.

[0072] S4: Use a trained multi-core single-class support vector machine to determine the sintering furnace operating conditions online and trigger an early warning.

[0073] In actual production, sensor data is generated continuously. The system needs to process the data stream in real time using the same window and feature extraction methods, and quickly provide a normal or abnormal judgment. Once an anomaly is detected, the system must issue an immediate warning to avoid quality accidents or equipment failures.

[0074] Temperature, pressure, and flow rate data inside the sintering furnace are acquired in real time at a sampling frequency of 100Hz, and a ring-shaped buffer with a length of 300 sampling points is maintained. Whenever a new sampling point is acquired, the buffer is updated, forming a new sliding window. The step size is 100 sampling points, meaning a new window is generated every 1 second.

[0075] Perform the same feature extraction process as S2 on the new window to obtain the multidimensional features of the new window.

[0076] Substitute the multidimensional features of the new window into the decision function of S3 to calculate the decision function value. If the decision function value is greater than or equal to 0, the new window is judged to be normal and no warning is issued; if it is less than 0, the new window is judged to be abnormal and an audible and visual alarm is immediately triggered, and the monitoring interface will display "Sintering furnace operation status deviates from normal mode" while recording the original data of the new window.

[0077] This completes the real-time online monitoring and early warning of the sintering furnace's operating conditions.

[0078] This invention also provides an intelligent early warning system for sintering furnaces of powder metallurgy parts. For example... Figure 3 As shown, the system includes a processor and a memory. The memory stores computer program instructions, which, when executed by the processor, implement the intelligent early warning method for sintering furnaces of powder metallurgy parts according to the first aspect of the present invention.

[0079] The system also includes other components well known to those skilled in the art, such as communication buses and communication interfaces, the settings and functions of which are known in the art and will not be described in detail here.

[0080] It should be noted that those skilled in the art can make various modifications and improvements without departing from the inventive concept, and these all fall within the scope of protection of this invention. Therefore, the scope of protection of this patent should be determined by the appended claims.

Claims

1. A method for intelligent early warning of sintering furnaces for powder metallurgy parts, characterized in that, include: Collect historical normal operation data of the sintering furnace and preprocess it to obtain a set of historical windows consisting of a sliding window of normal operating conditions; Multidimensional features of each window are extracted from the set of historical windows. The multidimensional features include: structural feature vectors extracted based on the singular values ​​of the autocorrelation matrix of each parameter, coupling feature vectors extracted based on the cross-correlation matrix between each parameter, and inertial-coupling dominant features and inertial-coupling consistency vectors extracted based on joint analysis of autocorrelation and cross-correlation. The extracted multidimensional features are used to form a training sample set. The multidimensional features of each training sample are divided into a first sub-vector and a second sub-vector. The first sub-vector is the structural feature vector, and the second sub-vector is a vector composed of features other than the structural feature vector. The kernel function corresponding to the first sub-vector is a cosine kernel, and the kernel function corresponding to the second sub-vector is a radial basis function kernel. The weights of the cosine kernel and the radial basis function kernel are determined according to the mean of the inertial-coupling dominant features of all training samples in the training sample set. The combined kernel function obtained by weighted summation is used to train a single-class support vector machine to obtain the decision function. The system collects real-time operating data of the sintering furnace, extracts multi-dimensional features from the real-time window, and substitutes them into the decision function of a pre-trained single-class support vector machine for discrimination. When the decision function value indicates an anomaly, an early warning is triggered.

2. The intelligent early warning method for sintering furnaces of powder metallurgy parts according to claim 1, characterized in that, The historical normal operation data includes time series data of at least three parameters: temperature, pressure, and flow rate. The preprocessing includes using a sliding window to divide the long time series into overlapping short time windows and normalizing the data in each window.

3. The intelligent early warning method for sintering furnaces of powder metallurgy parts according to claim 1, characterized in that, The structural feature vector is extracted in the following way: Select any window in the historical window set as the target window. For the target window, set the maximum lag order, calculate the autocorrelation coefficient of each parameter at each lag order and arrange them into an autocorrelation matrix. Perform singular value decomposition on the autocorrelation matrix to obtain multiple singular values. Concatenate the multiple singular values ​​in descending order to form a structural feature vector.

4. The intelligent early warning method for sintering furnaces of powder metallurgy parts according to claim 3, characterized in that, The method for extracting the coupling feature vector is as follows: For the target window, the same lag order is used. For each lag order, the cross-correlation function between any two parameters is calculated and arranged into a cross-correlation matrix to obtain the cross-correlation matrix for each lag order. For each directed parameter pair set according to historical normal operation data, the maximum value of the absolute value sequence of its cross-correlation values ​​under all lag orders is extracted, and the maximum values ​​corresponding to all directed parameter pairs are arranged in a fixed order to form a coupling feature vector.

5. The intelligent early warning method for sintering furnaces of powder metallurgy parts according to claim 4, characterized in that, The extraction method for the inertial-coupling dominant feature is as follows: For each lag, the mean of the absolute values ​​of the autocorrelation coefficients is calculated as the average inertial strength, and the mean of the absolute values ​​of all off-diagonal elements in the cross-correlation matrix is ​​calculated as the average coupling strength. Then, the inertial-coupling dominance ratio of the lag is obtained according to the relationship between the average inertial strength and the average coupling strength. The mean of the inertial-coupling dominance ratios of all lags is taken as the inertial-coupling dominance feature.

6. The intelligent early warning method for sintering furnaces of powder metallurgy parts according to claim 5, characterized in that, The inertial-coupling consistency vector is extracted as follows: For each parameter, the full-scale inertia vector corresponding to the parameter is extracted from the autocorrelation matrix, and the full-scale coupling vector corresponding to the parameter is extracted from the cross-correlation matrix group. The cosine similarity between the full-scale inertia vector and the full-scale coupling vector is calculated, and the cosine similarity of all parameters is concatenated to form the inertia-coupling consistency vector.

7. The intelligent early warning method for sintering furnaces of powder metallurgy parts according to claim 1, characterized in that, The weight of the cosine kernel is obtained by truncating the mean of the inertial-coupling dominant features of all training samples in the training sample set to a preset interval, and the weight of the radial basis kernel is equal to 1 minus the weight of the cosine kernel.

8. An intelligent early warning system for a sintering furnace of powder metallurgy parts, characterized in that, include: A processor and a memory, wherein the memory stores computer program instructions that, when executed by the processor, implement the intelligent early warning method for sintering furnaces of powder metallurgy parts according to any one of claims 1-7.

Images