Health monitoring method and device based on semi-mechanism model

Abstract

The invention provides a health monitoring method and device based on a semi-mechanism model. The method comprises the steps that the semi-mechanism model is used for converting machine equipment sensor signals into semi-mechanism features and extracting continuous numerical value features; performing feature discretization classification on the continuous numerical value features, and endowing real labels and real health degrees; using a health degree formula of undetermined parameters to generate a predicted health degree of all the continuous numerical value features; using the closeness degree of the predicted health degree of all the continuous numerical value features and the real health degree to obtain predicted labels of all the continuous numerical value features; taking minimization of the cross entropy as an optimization target, and using a parameter search method to determine undetermined parameters of the health degree formula; according to the continuous numerical value characteristics, box separation and the health degree of the box separation are adjusted in real time, and parameters of the health degree formula are updated. According to the method, health monitoring parameters can be dynamically adjusted to adapt to changes of sensor signals, and the health condition of the equipment can be diagnosed more accurately.

Description

Technical Field

[0001] The present invention relates to the technical field of industrial equipment monitoring, and in particular to a health monitoring method and device based on a semi-mechanism model. Background Art

[0002] Health monitoring of industrial equipment is of great significance for improving equipment reliability, optimizing production processes, improving product quality and achieving predictive maintenance. In the health monitoring of industrial equipment, the collected sensor signals usually appear in the form of time series data, which contains important information such as equipment operating status and changes in production process parameters. Due to the complexity of industrial equipment and changes in the surrounding environment, such as temperature changes, humidity changes, equipment load changes and working conditions, sensor signals often undergo unpredictable changes. Therefore, in order to dynamically adapt to changes in sensor signals and accurately diagnose the health status of equipment, adaptive health monitoring methods based on time series data are crucial.

[0003] In view of this, this application is filed. Summary of the invention

[0004] The present invention provides a health monitoring method and device based on a semi-mechanism model, which can at least partially improve the above-mentioned problems.

[0005] To achieve the above object, the present invention adopts the following technical solutions: A health monitoring method based on a semi-mechanism model, comprising: Acquire the sensor signal of the machine equipment to be identified collected by the sensor, use the preset semi-mechanism model to transform the sensor signal into a health formula to generate a semi-mechanism feature, and perform data preprocessing on the semi-mechanism feature to obtain a continuous numerical feature; The continuous numerical features are binned using a feature discretization method, and the continuous numerical features are classified according to the obtained bins to generate bin categories, and the bins of all bin categories are given real health degrees in combination with preset business knowledge and domain knowledge; The bin categories are encoded using a one-hot encoding method to generate a true label of a continuous numerical feature, a health formula with undetermined parameters is used to calculate a predicted health value of the continuous numerical feature, and the degree of proximity between the predicted health value and the true health degree is characterized to obtain a predicted label of the continuous numerical feature; The cross entropy is calculated based on the real label and the predicted label, and the minimization of the cross entropy is used as the optimization goal. In combination with the business knowledge and domain knowledge, the prior distribution of the undetermined parameters of the health formula is preset, and the optimal preset parameters of the health formula are determined according to the Bayesian hyperparameter search algorithm to obtain the final health formula. The final health formula is used to calculate the health of the continuous numerical feature to complete the health monitoring of the machine to be identified.

[0006] The present invention also provides a health monitoring device based on a semi-mechanism model, which comprises: An extraction unit is used to obtain a sensor signal of a machine device to be identified collected by a sensor, perform health formula conversion processing on the sensor signal using a preset semi-mechanism model to generate a semi-mechanism feature, and perform data preprocessing on the semi-mechanism feature to obtain a continuous numerical feature; An extraction unit, used to perform binning processing on the continuous numerical features using a feature discretization method, classify the continuous numerical features according to the obtained bins, generate bin categories, and assign true health to the bins of all bin categories in combination with preset business knowledge and domain knowledge; A processing unit, configured to encode the bin categories using a one-hot encoding method to generate a true label of a continuous numerical feature, calculate a predicted health value of the continuous numerical feature using a health formula of an undetermined parameter, and characterize the degree of proximity between the predicted health value and the true health, to obtain a predicted label of the continuous numerical feature; A monitoring unit is used to calculate the cross entropy according to the real label and the predicted label, to minimize the cross entropy as an optimization goal, to preset the prior distribution of the undetermined parameters of the health formula in combination with the business knowledge and domain knowledge, to determine the optimal preset parameters of the health formula according to the Bayesian hyperparameter search algorithm, to obtain a final health formula, and to use the final health formula to calculate the health of the continuous numerical feature to complete the health monitoring of the machine to be identified.

[0007] In summary, the health monitoring method based on the semi-mechanism model includes: using the semi-mechanism model to convert the sensor signal of the machine equipment into a semi-mechanism feature and extracting the continuous numerical feature; discretizing and classifying the continuous numerical feature, and assigning the real label and the real health; using the health formula of the undetermined parameters to generate the predicted health of all the continuous numerical features; using the closeness between the predicted health of all the continuous numerical features and the real health to obtain the predicted label of all the continuous numerical features; taking minimizing the cross entropy as the optimization goal, using the parameter search method to determine the undetermined parameters of the health formula; adjusting the bins and their health in real time for the continuous numerical features, and updating the parameters of the health formula. This method can dynamically adjust the health monitoring parameters to adapt to the changes in sensor signals and more accurately diagnose the health status of the equipment.

[0008] In short, the health monitoring method based on the semi-mechanism model is based on the semi-mechanism model, and uses cross entropy and hyperparameter search methods to provide a method for automatically adjusting parameters for equipment health monitoring. This method allows the health monitoring parameters to be dynamically adjusted according to different equipment conditions, equipment types and their status, so as to obtain more accurate equipment health monitoring and analysis results. BRIEF DESCRIPTION OF THE DRAWINGS

[0009] Figure 1 is a flow chart of a health monitoring method based on a semi-mechanism model provided in the first embodiment of the present invention; Figure 2 It is a module schematic diagram of a health monitoring device based on a semi-mechanism model provided by the second embodiment of the present invention. DETAILED DESCRIPTION

[0010] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0011] refer to Figure 1 As shown, the first embodiment of the present invention discloses a health monitoring method based on a semi-mechanism model, which can be executed by a health monitoring device based on a semi-mechanism model (hereinafter referred to as a monitoring device), and in particular, executed by one or more processors in the monitoring device to implement the following method: In this embodiment, the monitoring device may be a computer, a server, a workstation, a gateway or other device with data processing capabilities, and the present invention does not make any specific limitation.

[0012] S1, obtaining the sensor signal of the machine equipment to be identified collected by the sensor, using a preset semi-mechanism model to transform the sensor signal into a health formula to generate a semi-mechanism feature, and performing data preprocessing on the semi-mechanism feature to obtain a continuous numerical feature; Specifically, step S1 includes: obtaining a sensor signal of the machine device to be identified collected by a sensor with a preset sampling rate; The sensor signal is filtered using a filter suitable for the sensor, and the filtered sensor signal is converted into a semi-mechanism feature using a preset semi-mechanism model; Data preprocessing is performed on the semi-mechanistic features to extract continuous numerical features.

[0013] The semi-mechanistic features are subjected to data preprocessing to extract continuous numerical features, specifically: The semi-mechanism features are segmented using a sliding window, wherein adjacent sliding windows can overlap but cannot cover each other; Statistical features, frequency domain features, and time domain features within the sliding window are extracted as continuous numerical features.

[0014] In this embodiment, a vibration sensor is first installed at a suitable position of the machine equipment to be identified, and then a sensor sampling rate that is compatible with the scene requirements is set according to the current scene requirements. For example, the sampling rate of the vibration sensor can be set to 50,000 to collect the vibration time domain signal of the machine equipment.

[0015] In this embodiment, the sensor signal of the machine equipment is obtained, the sensor signal of the machine equipment is converted into a semi-mechanism feature using a preset semi-mechanism model, the semi-mechanism feature is preprocessed, and the continuous numerical feature is extracted. Specifically, the sensor signal of the machine equipment to be identified is acquired by a sensor with a set sampling rate; the sensor signal of the machine equipment is preprocessed using a mutually adaptive filter, the sensor signal of the machine equipment is converted into a semi-mechanism feature using a preset semi-mechanism model; the semi-mechanism feature is preprocessed, and the continuous numerical feature is extracted.

[0016] In this embodiment, the semi-mechanism features are segmented using a sliding window method. Adjacent sliding windows are allowed to overlap but cannot cover each other, thereby extracting statistical features, frequency domain features, time domain features, etc. within the sliding window as continuous numerical features. The length of the sliding window and the interval distance between adjacent sliding windows can be adjusted in a timely manner based on scenario requirements and domain knowledge.

[0017] Among them, the semi-mechanism features are segmented by using a fixed-length sliding window. For example, the sliding window length is 600, and the overlapping length of adjacent sliding windows is 400, so as to calculate the average value of all the semi-mechanism features in the sliding window as the continuous numerical feature. The length of the sliding window and the interval distance between adjacent sliding windows can be adjusted in time according to scene requirements and domain knowledge.

[0018] S2, binning the continuous numerical features using a feature discretization method, classifying the continuous numerical features according to the obtained bins, generating bin categories, and assigning true health to the bins of all bin categories in combination with preset business knowledge and domain knowledge; Specifically, step S2 includes: based on the distribution of the continuous numerical feature, using a feature discretization method to bin the continuous numerical feature to obtain a plurality of bins; Dividing the continuous numerical feature according to all bins to generate multiple bin categories; In combination with preset business knowledge and domain knowledge, a continuous value from 0 to 100 is used to assign a corresponding real health degree to each bin of the bin category.

[0019] Based on the distribution of the continuous numerical feature, the continuous numerical feature is binned using a feature discretization method, specifically: Performing deduplication processing on the continuous numerical features to obtain deduplicated continuous numerical features, sorting the deduplicated continuous numerical features in ascending order, and calculating their corresponding quantiles; Setting a quantile point according to the quantile, and performing binning processing on the continuous numerical feature based on the quantile point; The calculation formula of quantile is: , where Q(x) is the quantile, x is the continuous numerical feature, and rank(x) indicates the number of continuous numerical features that are less than or equal to the current continuous numerical feature in the deduplicated continuous numerical feature. , where N is the number of deduplicated continuous numerical features.

[0020] In this embodiment, according to the distribution of the continuous numerical features, the continuous numerical features are first binned using a feature discretization method; then the continuous numerical features are classified using all bins; finally, corresponding real health levels are assigned to all categories in combination with business and domain knowledge.

[0021] First, the continuous numerical features are deduplicated, and then the deduplicated continuous numerical features are arranged in ascending order and the corresponding quantiles are calculated. The quantile points are set in a quantile binning manner to bin the continuous numerical features. Specifically: Use the following formula to calculate the current continuous numerical feature Quantile : ,in, Indicates that the deduplicated continuous numerical feature is less than or equal to the current continuous numerical feature. The number of Indicates the number of deduplicated continuous numerical features; combined with scenario requirements and domain knowledge, select Quantile ,here , so get Boxes: , ,..., , ; Then the continuous numerical features are classified according to the range of the bins to achieve discretization. Combining the scenario requirements and domain knowledge, without loss of generality, a continuous value from 0 to 100 is used to assign healthiness to all bins, denoted as .

[0022] S3, encoding the bin categories using a one-hot encoding method to generate a true label of a continuous numerical feature, calculating a predicted health value of the continuous numerical feature using a health formula of an undetermined parameter, and characterizing the degree of proximity between the predicted health value and the true health, to obtain a predicted label of the continuous numerical feature; Specifically, step S3 includes: encoding the bin categories using a one-hot encoding method to generate true labels for all continuous numerical features, wherein the bin categories are converted into binary feature vectors based on the one-hot encoding method as true labels; Calculating the predicted health value of the continuous numerical feature using the health formula of the undetermined parameter, wherein the value range of the health formula of the undetermined parameter is a continuous numerical interval of [0, 100]; Characterize the degree of proximity between the predicted health value and the actual health value according to probability; Based on the degree of proximity, a predicted label of the continuous numerical feature is obtained.

[0023] The degree of proximity between the predicted health value and the actual health value is characterized according to probability, specifically: Use the probability value P between [0, 1] to represent the true health S and the predicted score The calculation formula of the probability value P is: ,in, represents the natural exponential function, w is the attenuation coefficient greater than 0, Represents the prediction score and The gap, q is a value between The tolerance error between .

[0024] In this embodiment, firstly, the one-hot encoding method is used to encode all the bin categories to generate the true labels of all the bins; then, the health formula of the undetermined parameters is used to generate the predicted health of all the continuous numerical features; then, the probability value is used to characterize the closeness between the predicted health of all the continuous numerical features and the true health; finally, the probability value is used to obtain the predicted labels of all the continuous numerical features.

[0025] Specifically, in this embodiment, the one-hot encoding method is used to convert all the bin categories into binary feature vectors as true labels. The value range of the health formula of the undetermined parameter is preset to a continuous numerical interval of [0, 100].

[0026] The predicted health of all continuous numerical features is The actual health , and to facilitate the subsequent calculation of cross entropy, without loss of generality, use The probability value between express and The degree of closeness between The closer to 0, the and The proximity is very low. The closer to 1, the and The closeness between them is very high. Assume is a value located in If the predicted health is Falling In addition, predicting health With the true score The degree of proximity is 0; if the prediction score Falling The following function is used to calculate the degree of proximity : ,in, represents the natural exponential function, represents an attenuation coefficient greater than 0, Represents the prediction score and The attenuation coefficient can be preset as a large positive number to ensure that When increasing, will also increase and approach 1 more quickly; using the probability value Replace the component 1 in the true label to obtain the predicted labels for all the continuous numerical features.

[0027] Specifically, for bins, that is, there are categories, perform one-hot encoding on the binned categories, and convert them into generated categories of length of A binary vector consisting of 0s and 1s is used as the true label of the bin. indivual; The health formula of the undetermined parameter can be preset as a function with a value range of [0, 100]. The function selection here needs to be combined with business and domain knowledge; for example, if the larger the amplitude of the continuous numerical feature, the lower the corresponding health, and the smaller the amplitude, the higher the corresponding health, then the following function can be used: ,in, represents the natural exponential function, and Indicates preset parameters that are pending.

[0028] S4, calculate the cross entropy according to the real label and the predicted label, take minimizing the cross entropy as the optimization goal, combine the business knowledge and domain knowledge, preset the prior distribution of the unknown parameters of the health formula, determine the optimal preset parameters of the health formula according to the Bayesian hyperparameter search algorithm, obtain the final health formula, and use the final health formula to calculate the health of the continuous numerical feature to complete the health monitoring of the machine to be identified.

[0029] Specifically, step S4 includes: calculating cross entropy according to the true label and the predicted label; Taking minimizing the cross entropy as the optimization goal, a hyperparameter search method is used to determine the undetermined parameters of the health formula; Combining the business knowledge and domain knowledge, multiple health alarm lines are set, different alarm events are confirmed, and the bins and their health are adjusted in real time to update the pending parameters of the health formula; The Bayesian hyperparameter search algorithm is used to determine the optimal preset parameters of the health formula and obtain the final health formula.

[0030] The calculation formula of the cross entropy with weight coefficient is: ,in, Indicates Boxes, Indicates The number of binned data, Indicates that it belongs to The weight coefficient of the cross entropy of the bins, Indicates The first The true label of each element, Representation and The corresponding predicted label.

[0031] In this embodiment, the cross entropy is first calculated using the real labels and predicted labels of all the continuous numerical features; then, minimizing the cross entropy is used as the optimization goal, and the hyperparameter search method is used to determine the pending parameters of the health formula; finally, several health alarm lines are set in combination with business and domain knowledge, and by confirming different alarm events, the bins and their health are adjusted in real time for the continuous numerical features, and the pending parameters of the health formula are updated.

[0032] To weaken the impact of sample imbalance, the cross entropy formula with weight coefficients is used: ,in, Indicates Boxes, Indicates The number of binned data is Indicates that it belongs to The weight coefficient of the cross entropy of the bins; Indicates The first The true label of the element, Representation and Corresponding prediction labels; Taking minimizing cross entropy as the optimization goal, presetting the prior distribution of the undetermined parameters of the health formula, and using the Bayesian hyperparameter search algorithm to determine the optimal preset parameters of the health formula. That is, using the hyperparameter search method to select the health formula parameters with the minimum cross entropy as the optimal parameter combination to determine the final health conversion formula.

[0033] Among them, the prior distribution of the undetermined parameters of the health formula is set, and the optimization goal is to minimize the cross entropy. The Bayesian hyperparameter search algorithm is used to determine the optimal preset parameters of the health formula. In combination with business and domain knowledge, the health alarm line is set, and the number of quantiles used for binning is adjusted by confirming the quantile range of the continuous numerical features corresponding to different alarm events, allocating the health corresponding to each bin, updating the undetermined coefficients of the health formula in real time, and redeploying equipment monitoring.

[0034] In summary, the health monitoring method based on the semi-mechanism model is based on the semi-mechanism model and uses cross entropy and hyperparameter search methods to provide a method for automatically adjusting parameters for equipment health monitoring. This method allows the health monitoring parameters to be dynamically adjusted according to different equipment conditions, equipment types and their states, so as to obtain more accurate equipment health monitoring and analysis results.

[0035] See also Figure 2 The second embodiment of the present invention provides a health monitoring device based on a semi-mechanism model, which includes: The extraction unit 201 is used to obtain the sensor signal of the machine equipment to be identified collected by the sensor, use the preset semi-mechanism model to transform the sensor signal into a health formula, generate a semi-mechanism feature, and perform data preprocessing on the semi-mechanism feature to obtain a continuous numerical feature; The labeling unit 202 is used to perform binning processing on the continuous numerical features using a feature discretization method, classify the continuous numerical features according to the obtained bins, generate bin categories, and assign real health degrees to the bins of all bin categories in combination with preset business knowledge and domain knowledge; The processing unit 203 is used to encode the bin categories by one-hot encoding to generate a true label of the continuous numerical feature, calculate the predicted health value of the continuous numerical feature by using the health formula of the undetermined parameter, and characterize the closeness between the predicted health value and the true health, so as to obtain the predicted label of the continuous numerical feature; The monitoring unit 204 is used to calculate the cross entropy according to the real label and the predicted label, to minimize the cross entropy as the optimization goal, to preset the prior distribution of the undetermined parameters of the health formula in combination with the business knowledge and domain knowledge, to determine the optimal preset parameters of the health formula according to the Bayesian hyperparameter search algorithm, to obtain the final health formula, and to use the final health formula to calculate the health of the continuous numerical feature to complete the health monitoring of the machine to be identified.

[0036] The above is a preferred embodiment of the present invention. It should be pointed out that a person skilled in the art can make several improvements and modifications without departing from the principle of the present invention. These improvements and modifications are also considered to be within the scope of protection of the present invention.

Claims

1. A health monitoring method based on a semi-mechanism model, characterized in that: include: Acquire the sensor signal of the machine equipment to be identified collected by the sensor, use the preset semi-mechanism model to transform the sensor signal into a health formula to generate a semi-mechanism feature, and perform data preprocessing on the semi-mechanism feature to obtain a continuous numerical feature; The continuous numerical features are binned using a feature discretization method, and the continuous numerical features are classified according to the obtained bins to generate bin categories, and the bins of all bin categories are given real health degrees in combination with preset business knowledge and domain knowledge; The bin categories are encoded using a one-hot encoding method to generate a true label of a continuous numerical feature, a health formula with undetermined parameters is used to calculate a predicted health value of the continuous numerical feature, and the degree of proximity between the predicted health value and the true health degree is characterized to obtain a predicted label of the continuous numerical feature; The cross entropy is calculated based on the real label and the predicted label, and the minimization of the cross entropy is used as the optimization goal. In combination with the business knowledge and domain knowledge, the prior distribution of the undetermined parameters of the health formula is preset, and the optimal preset parameters of the health formula are determined according to the Bayesian hyperparameter search algorithm to obtain the final health formula. The final health formula is used to calculate the health of the continuous numerical feature to complete the health monitoring of the machine to be identified.

2. The health monitoring method based on the semi-mechanism model according to claim 1 is characterized in that: The sensor signal of the machine to be identified is acquired by the sensor, and the sensor signal is transformed into a health formula using a preset semi-mechanism model to generate a semi-mechanism feature, and the semi-mechanism feature is preprocessed to obtain a continuous numerical feature, specifically: Obtaining sensor signals of the machine equipment to be identified collected by a sensor with a preset sampling rate; The sensor signal is filtered using a filter suitable for the sensor, and the filtered sensor signal is converted into a semi-mechanism feature using a preset semi-mechanism model; Data preprocessing is performed on the semi-mechanistic features to extract continuous numerical features.

3. The health monitoring method based on the semi-mechanism model according to claim 2 is characterized in that: The semi-mechanistic features are subjected to data preprocessing to extract continuous numerical features, specifically: The semi-mechanism features are segmented using a sliding window, wherein adjacent sliding windows can overlap but cannot cover each other; Statistical features, frequency domain features, and time domain features within the sliding window are extracted as continuous numerical features.

4. The health monitoring method based on a semi-mechanism model according to claim 1, characterized in that: The continuous numerical features are binned using a feature discretization method, and the continuous numerical features are classified according to the obtained bins to generate bin categories. The bins of all bin categories are given real health degrees in combination with preset business knowledge and domain knowledge, specifically: Based on the distribution of the continuous numerical feature, the continuous numerical feature is binned using a feature discretization method to obtain a plurality of bins; Dividing the continuous numerical feature according to all bins to generate multiple bin categories; In combination with preset business knowledge and domain knowledge, a continuous value from 0 to 100 is used to assign a corresponding real health degree to each bin of the bin category.

5. The health monitoring method based on the semi-mechanism model according to claim 4 is characterized in that: Based on the distribution of the continuous numerical feature, the continuous numerical feature is binned using a feature discretization method, specifically: Performing deduplication processing on the continuous numerical features to obtain deduplicated continuous numerical features, sorting the deduplicated continuous numerical features in ascending order, and calculating their corresponding quantiles; Setting a quantile point according to the quantile, and performing binning processing on the continuous numerical feature based on the quantile point; The calculation formula of quantile is: , where Q(x) is the quantile, x is the continuous numerical feature, and rank(x) indicates the number of continuous numerical features that are less than or equal to the current continuous numerical feature in the deduplicated continuous numerical feature. , where N is the number of deduplicated continuous numerical features.

6. The health monitoring method based on a semi-mechanism model according to claim 1, characterized in that: The binning categories are encoded using a one-hot encoding method to generate the true label of the continuous numerical feature, the predicted health value of the continuous numerical feature is calculated using the health formula of the undetermined parameter, and the degree of proximity between the predicted health value and the true health is characterized to obtain the predicted label of the continuous numerical feature, specifically: Encoding the bin categories using a one-hot encoding method to generate true labels for all continuous numerical features, wherein the bin categories are converted into binary feature vectors based on the one-hot encoding method as true labels; Calculating the predicted health value of the continuous numerical feature using the health formula of the undetermined parameter, wherein the value range of the health formula of the undetermined parameter is a continuous numerical interval of [0, 100]; Characterize the degree of proximity between the predicted health value and the actual health value according to probability; Based on the degree of proximity, a predicted label of the continuous numerical feature is obtained.

7. The health monitoring method based on the semi-mechanistic model according to claim 6 is characterized in that: The degree of proximity between the predicted health value and the actual health value is characterized according to probability, specifically: Use the probability value P between [0, 1] to represent the true health S and the predicted score The calculation formula of the probability value P is: ,in, represents the natural exponential function, w is the attenuation coefficient greater than 0, Represents the prediction score and The gap, q is a value between The tolerance error between .

8. The health monitoring method based on a semi-mechanistic model according to claim 1, characterized in that: The cross entropy is calculated according to the real label and the predicted label, and the minimization of the cross entropy is used as the optimization goal. The prior distribution of the undetermined parameters of the health formula is preset in combination with the business knowledge and domain knowledge. The optimal preset parameters of the health formula are determined according to the Bayesian hyperparameter search algorithm to obtain the final health formula, which is specifically, Calculate cross entropy based on the true label and the predicted label; Taking minimizing the cross entropy as the optimization goal, a hyperparameter search method is used to determine the undetermined parameters of the health formula; Combining the business knowledge and domain knowledge, multiple health alarm lines are set, different alarm events are confirmed, and the bins and their health are adjusted in real time to update the pending parameters of the health formula; The Bayesian hyperparameter search algorithm is used to determine the optimal preset parameters of the health formula and obtain the final health formula.

9. The health monitoring method based on the semi-mechanistic model according to claim 8 is characterized in that: The calculation formula of the cross entropy with weight coefficient is: ,in, Indicates Boxes, Indicates The number of binned data, Indicates that it belongs to The weight coefficient of the cross entropy of the bins, Indicates The first The true label of each element, Representation and The corresponding predicted label.

10. A health monitoring device based on a semi-mechanism model, characterized in that: include: An extraction unit is used to obtain a sensor signal of a machine device to be identified collected by a sensor, perform health formula conversion processing on the sensor signal using a preset semi-mechanism model to generate a semi-mechanism feature, and perform data preprocessing on the semi-mechanism feature to obtain a continuous numerical feature; A labeling unit, used to perform binning processing on the continuous numerical features using a feature discretization method, classify the continuous numerical features according to the obtained bins, generate bin categories, and assign true health to the bins of all bin categories in combination with preset business knowledge and domain knowledge; A processing unit, configured to encode the bin categories using a one-hot encoding method to generate a true label of a continuous numerical feature, calculate a predicted health value of the continuous numerical feature using a health formula of an undetermined parameter, and characterize the degree of proximity between the predicted health value and the true health, to obtain a predicted label of the continuous numerical feature; A monitoring unit is used to calculate the cross entropy according to the real label and the predicted label, to minimize the cross entropy as an optimization goal, to preset the prior distribution of the undetermined parameters of the health formula in combination with the business knowledge and domain knowledge, to determine the optimal preset parameters of the health formula according to the Bayesian hyperparameter search algorithm, to obtain a final health formula, and to use the final health formula to calculate the health of the continuous numerical feature to complete the health monitoring of the machine to be identified.

Images