Safety hook heat treatment process performance evaluation method and device based on neural network

Abstract

The invention relates to the technical field of process monitoring, in particular to a safety hook heat treatment process performance evaluation method and device based on a neural network. The method comprises the following steps: in a process of processing a safety hook by a to-be-evaluated heat treatment process, collecting various to-be-evaluated process data of the to-be-evaluated heat treatment process; inputting the multiple kinds of to-be-evaluated process data into a pre-trained data compression model to obtain to-be-evaluated compressed data corresponding to the multiple kinds of to-be-evaluated process data; inputting the to-be-evaluated compression data into a pre-trained heat treatment process performance evaluation model to obtain an evaluation result of the to-be-evaluated heat treatment process; according to the invention, the problem of low evaluation quality of a manual evaluation mode in the prior art can be solved.

Description

Technical Field

[0001] The present application relates to the field of artificial intelligence technology, and in particular, to a method and device for evaluating the heat treatment process performance of safety hooks based on neural networks. Background Art

[0002] In production and life, safety hooks are stress tools or stress components widely used in infrastructure industries such as construction, mainly used to realize the movement, tying, and stabilization of spatial structures of objects. Safety hooks are widely used in industries and fields such as aviation, aerospace, ports, docks, logistics, mine traction and lifting, and offshore platforms. In the manufacturing process of safety hooks, the heat treatment process is a very important key link, and its process quality directly affects the mechanical properties, service life, and safety of safety hooks. Therefore, it is necessary to evaluate the heat treatment process of safety hooks to determine whether the obtained safety hooks meet the usage requirements.

[0003] In the prior art, the heat treatment process of safety hooks is usually evaluated by manual evaluation. However, since the heat treatment process of safety hooks involves multiple parameters, such as temperature, pressure, atmosphere composition, applied force, and material hardness, and in addition, due to the complex non-linear relationship between the above parameters, the manual evaluation method has the problem of low evaluation quality. Summary of the Invention

[0004] In view of this, the purpose of the present application is to provide a method and device for evaluating the heat treatment process performance of safety hooks based on neural networks to solve the problem of low evaluation quality in the manual evaluation method existing in the prior art.

[0005] In a first aspect, the present application provides a method for evaluating the heat treatment process performance of safety hooks based on neural networks, and the method includes: During the process of treating the safety hook with the heat treatment process to be evaluated, collect a variety of process data to be evaluated of the heat treatment process to be evaluated; Input the variety of process data to be evaluated into a pre-trained data compression model to obtain compressed data to be evaluated corresponding to the variety of process data to be evaluated; Input the compressed data to be evaluated into a pre-trained heat treatment process performance evaluation model to obtain an evaluation result of the heat treatment process to be evaluated.

[0006] In a second aspect, the present application provides a device for evaluating the heat treatment process performance of safety hooks based on neural networks, and the device includes: a collection module, a compression module, and an evaluation module; The collection module is used to collect a variety of process data to be evaluated of the heat treatment process to be evaluated during the process of treating the safety hook with the heat treatment process to be evaluated; The compression module is configured to input a variety of the process data to be evaluated into a pre-trained data compression model to obtain the compression data to be evaluated corresponding to the variety of the process data to be evaluated; The evaluation module is configured to input the compression data to be evaluated into a pre-trained heat treatment process performance evaluation model to obtain the evaluation result of the heat treatment process to be evaluated.

[0007] Beneficial effects of the neural network: The present application provides a method for evaluating the performance of a safety hook heat treatment process based on a neural network. The method includes: during the process of treating a safety hook with a heat treatment process to be evaluated, collecting a variety of process data to be evaluated of the heat treatment process to be evaluated; inputting the variety of process data to be evaluated into a pre-trained data compression model to obtain the compression data to be evaluated corresponding to the variety of process data to be evaluated; inputting the compression data to be evaluated into a pre-trained heat treatment process performance evaluation model to obtain the evaluation result of the heat treatment process to be evaluated; in summary, it can be seen that since the method for evaluating the performance of a safety hook heat treatment process based on a neural network provided by the present application can realize the automatic evaluation of the heat treatment process of the safety hook and avoid manual evaluation. In addition, the process data to be evaluated is compressed during the automatic evaluation process, ensuring the accuracy of the evaluation. Therefore, it can solve the problem that the evaluation method of manual evaluation in the prior art has a low evaluation quality. Description of the Drawings

[0008] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings required to be used in the embodiments of the present application will be briefly introduced below. The following drawings only show some embodiments of the present application, and therefore should not be regarded as limiting the scope. For those of ordinary skill in the art, other related drawings can also be obtained based on these drawings without creative efforts.

[0009] Figure 1 It is a schematic flowchart of the method for evaluating the performance of a safety hook heat treatment process based on a neural network provided by an embodiment of the present application; Figure 2 It is an example diagram of the comparison of training losses of different training methods provided by an embodiment of the present application; Figure 3 It is an example diagram of the influence of the autoencoder on the reconstruction accuracy provided by an embodiment of the present application; Figure 4 It is an example diagram of the influence of gradient conflict detection on parameter update provided by an embodiment of the present application; Figure 5 It is an example diagram of the comparison of the energy flow on the feature retention ability in a noise environment provided by an embodiment of the present application; Figure 6Schematic structural diagram of an evaluation device for the heat treatment process performance of a safety hook based on a neural network provided by an embodiment of the present application. Detailed implementation manners

[0010] To make the objectives, technical solutions, and advantages of the embodiments of the present application clearer, the technical solutions of the present application will be clearly and completely described below with reference to the accompanying drawings. Apparently, the described embodiments are some, but not all, of the embodiments of the present application. All other embodiments obtained by those of ordinary skill in the art based on the embodiments in the present application without creative efforts shall fall within the protection scope of the present application.

[0011] First, the present application provides a method for evaluating the heat treatment process performance of a safety hook based on a neural network, as Figure 1 shown Figure 1 is a schematic flowchart of the method for evaluating the heat treatment process performance of a safety hook based on a neural network provided by an embodiment of the present application. The method includes: S110 - S130, details are as follows: S110: During the process of treating the safety hook with the heat treatment process to be evaluated, collect various process data to be evaluated of the heat treatment process to be evaluated.

[0012] Specifically, compared with the evaluation method by manual evaluation in the prior art, the technical solution of the present application can automatically evaluate the heat treatment process to be evaluated through the evaluation device for the heat treatment process performance of the safety hook, avoiding the problem of low evaluation quality caused by relying on manual evaluation.

[0013] In actual operation, the sources of the process data to be evaluated are mainly of two types; among them, the first is to directly collect at devices such as the temperature control system and mechanical testing equipment of the heat treatment furnace required for the heat treatment process, and the collected data may include control system logs, mechanical operation records, and relevant experimental data, etc.; the second is obtained by collecting through multiple sensors arranged at the site of the heat treatment process to be evaluated, and the sensors at least include, such as, temperature sensors, pressure sensors, and stress-strain sensors, etc.

[0014] In actual operation, the types of process data to be evaluated include: furnace temperature (unit: °C, representing the temperature inside the furnace during heat treatment), atmosphere temperature (unit: °C, representing the atmosphere temperature inside the furnace), applied force (unit: N, representing the force applied during heat treatment), heat treatment time (unit: seconds, representing the duration of a certain heat treatment cycle), material type identifier (representing the type of material being processed, such as steel, aluminum, etc.), material hardness (unit: HRC, representing the hardness of the material after heat treatment), strain rate of the material (unit: %, representing the deformation of the material during heat treatment), concentration of the furnace atmosphere (unit: ppm, representing the concentration of chemical components in the furnace atmosphere, such as oxygen, nitrogen, etc.), temperature gradient (unit: °C / min, representing the rate of temperature change inside the furnace), etc.; among them, it should be emphasized that the above data to be evaluated should all carry timestamps for integrating the data corresponding to the same timestamp into one input data.

[0015] During the process of collecting the process data to be evaluated, the real-time data of each sensor can be aggregated to the background server through serial or parallel transmission protocols (such as Modbus, CAN-bus) for enabling the background server to execute S120~S130; among them, the reading accuracy, data volume, and sampling frequency of the sensor are all dynamically adjusted by the background server to meet the requirements under different operating conditions.

[0016] S120: Input various process data to be evaluated into a pre-trained data compression model to obtain the compressed data to be evaluated corresponding to various process data to be evaluated.

[0017] Specifically, in actual application, the process data to be evaluated includes hundreds of data that can be collected by sensors, such as furnace temperature, applied force, and material hardness. If the above hundreds of data are all used as input data and input into the heat treatment process performance evaluation model, it will lead to an extremely high feature dimension of the input data input into the heat treatment process performance evaluation model. High-dimensional data is prone to cause the curse of dimensionality, increase the calculation difficulty of the model, and further reduce the calculation efficiency of the model. In addition, the sparse data distribution may weaken the generalization ability of the model; thus, it can be seen that if the above hundreds of data are all used as input data and input into the heat treatment process performance evaluation model for evaluating the heat treatment process to be evaluated, it may lead to the heat treatment process performance evaluation model being unable to make accurate evaluation results.

[0018] To solve this problem, this application sets up a data compression model for compressing the process data to be evaluated collected by multiple sensors, and inputs the compressed data to be evaluated corresponding to various process data to be evaluated output by the data compression model into the heat treatment process performance evaluation model to obtain accurate evaluation results.

[0019] In one implementation, the network structure of the data compression model is an autoencoder, and the autoencoder includes an encoder and a decoder; in the process of iteratively training the autoencoder, the current training process before reaching the training stop condition includes steps (1) to (3), details are as follows: Step (1): According to multiple training samples, determine the current sample input data corresponding to each training sample.

[0020] Among them, the training samples include various sample process data, and the sample process data is collected during the sample heat treatment process of the safety hook.

[0021] Specifically, in the embodiment of the present application, the task of the encoder in the trained autoencoder is to compress the process data to be evaluated from a high-dimensional space to a low-dimensional latent space, and the decoder attempts to reconstruct the process data to be evaluated from the low-dimensional representation.

[0022] Before using the autoencoder to determine the compressed data to be evaluated corresponding to multiple process data to be evaluated, it is necessary to iteratively train the autoencoder; in actual operation, in order to adapt to the complexity of the process data to be evaluated, for the training of the autoencoder, on the basis of using the gradient descent method for parameters, an inertial calibration method is used to optimize the gradient update step size of each round of training process. The inertial calibration method ensures smoother and more stable parameter updates by incorporating the influence of historical gradients into the parameter update process, and avoids the oscillation and convergence problems in traditional gradient descent.

[0023] It should be emphasized that during the iterative training of the autoencoder, the sample input data used in each training process should be the same. However, in order to distinguish different training rounds, the sample input data corresponding to different training rounds is distinguished in the embodiment of the present application. For example, "current sample input data" is used to represent the sample input data input to the autoencoder in the current training process, and "historical sample input data" is used to represent the sample input data input to the autoencoder in the historical training process before the current training process in terms of time sequence.

[0024] Step (2): Input the current sample input data into the encoder to obtain the current low-dimensional latent space representation feature corresponding to the current sample input data.

[0025] Specifically, in the embodiment of the present application, the parameters of both the encoder and the decoder include weight matrices; among them, the mapping method of the encoder is as follows: ; In the formula, represents the current low-dimensional latent space representation feature output by the encoder in the th training process, including Dimensional feature; represents an encoder function, characterizing the process of mapping the current sample input data to a low-dimensional latent space; represents the current sample input data input to the encoder during the represents the current weight matrix used by the encoder during the training process, which is a training parameter; represents the bias used by the encoder during the training process;

[0026] Specifically, the mapping method of the decoder is as follows: ; In the formula, represents the current sample compressed data output by the decoder during the training process; represents a decoder function, characterizing the process of recovering from the current low-dimensional latent space representation feature to the current sample compressed data; represents the weight matrix used by the decoder during the training process, which is the transpose of ; represents the bias used by the decoder during the training process, which is the transpose of

[0027] In actual operation, during the current training process before reaching the training stop condition of the autoencoder, after obtaining the current sample compressed data, it is also necessary to determine indicators such as the prediction accuracy of the autoencoder during the current training process according to the current sample input data and / or the current sample compressed data. In addition, it is also necessary to determine the parameters required by the autoencoder during the training process according to the current sample input data and the current sample compressed data.

[0028] It should be emphasized that the training stop condition of the autoencoder can be determined according to actual needs, and this application does not make specific limitations in this regard; in actual operation, the training stop condition can be to reach a preset maximum number of iterations. Preferably, the preset maximum number of iterations is set to 500 times, that is, when the training of the autoencoder reaches 500 times, the training can be stopped to obtain the sample compression model.

[0029] In one implementation, before step (1), S110 further includes steps (4) to (8), details are as follows: Step (4): Determine the historical loss function value of the autoencoder during the historical training process according to the historical latent space representation features output by the encoder during the historical training process.

[0030] In one implementation, step (4) includes steps (4.1) to (4.2), details are as follows: Step (4.1): Determine the feature weight weighting coefficient, the energy flow loss function value, and the adaptive feature interaction optimization loss function value according to the historical latent space representation features; Among them, the feature weight weighting coefficient indicates the influence of the historical latent space representation features on the target increment; the energy flow loss function value indicates the influence of the energy flow of the historical latent space representation features on the target increment, and the energy flow indicates the information amount of the historical latent space representation features; the adaptive feature interaction optimization loss function value indicates the influence of the features of each dimension in the historical latent space representation features on the target increment.

[0031] In one implementation, step (4.1) includes steps (4.1.1) to (4.1.2), details are as follows: Step (4.1.1): Determine the feature contribution degree and information gain of the historical latent space representation features according to the historical latent space representation features; Among them, the feature contribution degree indicates the contribution of the historical latent space representation features to the historical sample compression data output by the decoder during the historical training process; the information gain indicates the information amount difference of the historical latent space representation features.

[0032] Specifically, the feature contribution degree of the latent space representation features is calculated by considering the matching degree between the dimension of the process data to be evaluated and the overall distribution of the process data to be evaluated, and combining the variance of the feature and the entropy value of the feature; for example, the variance of a certain stress-strain sensor data is high but the entropy value is low (the distribution is concentrated), indicating that it has a significant and stable process influence and a high contribution degree, while the experimental parameters with high entropy (such as discrete debugging records) may be suppressed. Therefore, the feature contribution degree of the latent space representation features can balance between redundant feature (such as repeated log entries) compression and key feature (such as mutated temperature data) expansion, and improve the dimensionality reduction efficiency.

[0033] In one implementation, step (4.1.1) includes steps (4.1.1.1) to (4.1.1.3), details are as follows: Step (4.1.1.1): Determine the variance and entropy of the low-dimensional latent space representation features based on the historical latent space representation features; where the entropy indicates the degree of dispersion of the low-dimensional latent space representation features.

[0034] Specifically, the entropy in the space representation ensures that features with a large degree of variation but high uncertainty can be appropriately compressed during dimensional expansion; where the entropy is calculated as follows: ; In the formula, represents the probability distribution of the th candidate value of the low-dimensional latent space representation feature determined according to the th training process; represents the logarithmic function with base 10.

[0035] Step (4.1.1.2): Determine the feature contribution degree according to the preset second adjustment coefficient, historical latent space representation features, variance and entropy; where the second adjustment coefficient is used to adjust the relative relationship between the variance and the entropy.

[0036] Specifically, the formula for calculating the feature contribution degree is as follows: ; In the formula, represents the preset second adjustment coefficient; represents the entropy of the low-dimensional latent space representation feature determined according to the th training process.

[0037] Step (4.1.1.3): Determine the information gain according to the variance.

[0038] Specifically, the information gain of the latent space representation features automatically selects and enhances the representation of important features according to the feature distribution information of the low-dimensional latent space, and compresses redundant features at the same time. The information gain is used to select the features with the maximum amount of information.

[0039] For example, in the process data to be evaluated, for a multi-source sensor fusion scenario (such as simultaneously monitoring temperature, pressure, and strain), the data of multiple temperature sensors may show different distribution characteristics. Some sensors are located at key process nodes (such as the material heating area), and their temperature fluctuation variances are large, reflecting the core changes of the process. Other redundant sensors (such as ambient temperature monitoring) may have smaller variances and repeated information. Through the information gain Through calculation, the autoencoder can quantify the information content difference of each sensor feature, automatically expand the potential space dimension of high-gain features (such as key temperature nodes), while compressing the redundant features of low gain (such as ambient temperature), avoiding information dilution caused by feature dimension expansion, and ensuring that high-value sensor data can still dominate the representation of the potential space after dimensionality reduction.

[0040] Among them, the information gain is calculated as follows: ; In the formula, represents the variance of the feature of the th dimension in the low-dimensional potential space representation feature output by the encoder during the th training process.

[0041] Step (4.1.2): Determine the feature weight weighting coefficient according to the preset first adjustment coefficient, historical potential space representation features, feature contribution degree, and information gain; Among them, the first adjustment coefficient is used to adjust the compression and reconstruction of the historical sample input data by the encoder and decoder.

[0042] Specifically, different from the traditional autoencoder with a fixed potential dimension that cannot be dynamically adjusted according to the data distribution, the feature weight weighting coefficient characterizes the dynamic dimension expansion and contraction constraint of the feature. By calculating the variance of the potential space representation feature, the contribution degree of the feature is judged, and the potential space dimension is adjusted accordingly.

[0043] For example, in the process data to be evaluated, if the variance of a certain temperature sensor feature is significant (reflecting the key temperature changes in the process), the dimension of its corresponding low-dimensional potential space will be expanded to enhance the representation ability, while redundant mechanical operation records (such as repetitive log entries) will be compressed.

[0044] Among them, the calculation formula of the feature weight weighting coefficient is as follows: ; In the formula, represents the exponential function; represents the preset first adjustment coefficient; represents the th training process variance of the low-dimensional potential space representation feature output by the encoder; represents the low-dimensional potential space representation feature determined according to the th training process Feature contribution degree; Indicates the low-dimensional latent space representation feature determined according to the information gain.

[0045] In one implementation, step (4.1) further includes: step (4.1.3) to step (4.1.4), details are as follows: Step (4.1.3): Determine the energy flow of the low-dimensional latent space representation feature and the deviation degree of the energy flow according to the historical latent space representation feature.

[0046] Step (4.1.4): Determine the energy flow loss function value according to the preset third adjustment coefficient, energy flow and deviation degree; wherein, the third adjustment coefficient is used to adjust the relative relationship between the energy flow and entropy.

[0047] Specifically, in actual operation, the mechanical operation records and multi-sensor data in the process data to be evaluated may have redundancy (such as duplicate log entries) or noise (such as abnormal device start-stop signals). Redundant features waste computing resources, and noise interferes with the model's capture of key process parameters. That is, when the dimension of the process data to be evaluated is relatively high, conventional dimension expansion and compression methods are difficult to effectively balance information retention and feature selectivity; in addition, features with lower entropy have more concentrated information and stronger energy flow; while features with higher entropy have weaker energy flow due to their dispersed information. For example, the energy flow weight of high-entropy pressure sensor data (with a dispersed distribution) is reduced, while higher energy is given to low-entropy stable experimental parameters (such as a constant stress value), so as to achieve selective retention of features.

[0048] To solve the above problems, the embodiment of the present application adopts a feature balance optimization method based on energy flow, dynamically adjusts the importance of features during the feature compression process, so as to achieve efficient control of information volume and redundancy; the feature balance optimization method based on energy flow is inspired by the principle of energy conservation, quantifies and dynamically allocates the energy of each feature through a feature flow function, so that the feature information flow during the training process flows in a more reasonable way throughout the network. By optimizing the information flow between features, it is ensured that the energy flow intensity of important features is relatively large, while the energy flow of redundant features is effectively suppressed.

[0049] The energy flow of a feature not only considers the amplitude of the feature (measured by ), but also considers the entropy value of the feature to adjust the sparsity of the information volume. The energy flow loss function constructed according to the principle of dynamic energy balance in the embodiment of the present application dynamically adjusts the selection and compression process of features according to the energy flow of features; during the compression process, the selection of features not only depends on variance and entropy, but is also dynamically affected by the energy flow of features.

[0050] For example, high energy flow is assigned to the pressure sensor data with high-frequency changes to retain details, while the energy weight is reduced for the mechanical operation records with low information content (such as periodic logs). When processing multi-sensor fusion data (such as the joint monitoring of temperature, pressure, and strain), it is possible to avoid a single feature dominating the latent space and ensure an equilibrium representation of multi-dimensional information.

[0051] Among them, the value of the energy flow loss function is calculated as follows: ; ; ; In the formula, represents the energy flow of the low-dimensional latent space representation feature determined according to the -th training process; represents a preset third adjustment coefficient. In actual operation, can be set to 0.2; represents the deviation degree of the energy flow determined according to the -th training process; represents a preset reference energy flow. In actual operation, can be set to 10.

[0052] In one implementation, step (4.1) further includes: steps (4.1.5) to (4.1.7), which are as follows: Step (4.1.5): Determine the feature interaction matrix according to the historical sample input data input to the encoder during the historical training process; Among them, the feature interaction matrix indicates the interaction result between multiple features included in the historical sample input data.

[0053] Specifically, in actual operation, there are complex interactions among various feature data included in the process data to be evaluated (such as the non-linear coupling of temperature and stress affecting material deformation), and traditional linear dimensionality reduction (such as the principal component analysis method) cannot model high-order interactions, resulting in the loss of key process relationships.

[0054] To solve this problem, the embodiments of the present application solve this problem through the method of adaptive feature interaction optimization; the feature balance optimization method based on energy flow can adjust the intensity of features, but when the process data to be evaluated contains complex non-linear relationships, there is still a possibility that high-order interactions between some features are ignored. The embodiments of the present application adopt an adaptive feature interaction optimization mechanism to automatically identify and optimize the high-order interaction relationships between features, so as to enhance the model's ability to model non-linear and complex feature relationships.

[0055] Traditional compression methods such as principal component analysis and autoencoders usually ignore the high-order interactions between features during compression. For example, the quadratic interaction between stress data and temperature data may reflect the thermal expansion effect of materials, while the cubic term can describe the non-linear deformation characteristics. To better capture these interactions, the adaptive feature interaction optimization mechanism constructs a feature interaction matrix to describe the second-order or higher-order interactions between features, enhancing the modeling ability for complex process relationships (such as the mechanical property changes under the cooperation of multiple sensors), and making up for the deficiencies of traditional linear dimensionality reduction methods. Suppose there are features in the input process data set to be evaluated by the encoder.

[0056] Among them, the feature interaction matrix is calculated as follows: ; In the formula, represents the th power of the sample input data input to the encoder during the th training process. The high-order power can be used to capture its high-order interaction effect.

[0057] Step (4.1.6): Perform normalization processing on the feature interaction matrix to obtain the normalized feature interaction matrix.

[0058] Specifically, in the embodiment of the present application, the feature interaction matrix is normalized to ensure that features of different orders can have balanced weights during the training process. The normalization method constructs a matrix based on the high-order power of the input data and performs L2 norm normalization to balance the weights of features of different orders and enhance the modeling ability for complex non-linear relationships.

[0059] Among them, the normalized feature interaction matrix is calculated as follows: ; In the formula, represents the L2 norm of the feature interaction matrix.

[0060] Step (4.1.7): Determine the adaptive feature interaction optimization loss function value according to the preset weight coefficient, the regularization coefficient of the preset adaptive feature interaction optimization loss function value, the low-dimensional latent space representation feature, and the historical weight matrix; Among them, the weight coefficient adjusts the influence of the low-dimensional latent space representation feature on the autoencoder; the regularization coefficient is used to prevent overfitting.

[0061] Specifically, in the embodiments of the present application, in order to enable the model to adaptively learn the high-order interaction relationships between features, an adaptive feature interaction optimization loss function is adopted. , based on the contribution degree of each feature to the final prediction result, automatically adjusts the weight of its interaction relationship; the adaptive feature interaction optimization loss function functions to constrain the autoencoder to not only minimize the prediction error, but also optimize the interaction relationship between features by adjusting the feature weights in the high-order interaction matrix, thereby effectively capturing complex non-linear relationships.

[0062] Among them, the value of the adaptive feature interaction optimization loss function is calculated as follows: ; In the formula, represents the weight coefficient of the preset low-dimensional latent space representation feature , which is used to control the influence of the latent space representation feature on the autoencoder. In actual operation, can be set to 0.2; represents the normalized feature interaction matrix determined according to the th training process; represents the label of the training sample; represents the preset value of the adaptive feature interaction optimization loss function 's regularization coefficient, which is used to prevent overfitting; represents the feature interaction matrix determined according to the th training process; represents the Softmax function.

[0063] Step (4.2): Determine the historical loss function value according to the historical sample input data input to the encoder during the historical training process, the preset sparsity weighting coefficient, the feature weight weighting coefficient, the energy flow loss function value, and the adaptive feature interaction optimization loss function value.

[0064] Specifically, in the feature reconstruction stage, through the dynamic dimension expansion and contraction method, the dimension of the latent space is gradually optimized. Different from the traditional autoencoder that simply reconstructs in the low-dimensional latent space, the dynamic dimension expansion and contraction method adopted in the embodiments of the present application can adaptively adjust the dimension of the low-dimensional latent space according to the characteristics of the process data to be evaluated. When the feature dimension is high or unevenly distributed, the decoder will dynamically expand the dimension of the low-dimensional latent space according to the contribution degree of each feature, while redundant or irrelevant features will be compressed.

[0065] In the embodiments of the present application, different from traditional loss functions that only consider reconstruction errors, the loss function of the autoencoder is calculated based on the norm of the reconstructed data, the energy flow loss, and the feature interaction loss, suppressing redundancy and capturing high-order interaction relationships while reconstructing the data.

[0066] Among them, the value of the loss function of the autoencoder is calculated as follows: ; In the formula, represents the dimension of the features included in the sample input data; represents the feature weight weighting coefficient determined according to the th training process; represents a preset sparsity weighting coefficient. In actual operation, can be set to 0.3.

[0067] represents the L2 norm, the same as the calculation method of the Euclidean distance; represents the L1 norm.

[0068] represents the th dimension data in the sample input data input to the encoder during the th training process; represents the th dimension data in the sample compressed data output by the decoder during the th training process; represents the value of the energy flow loss function determined according to the th training process; represents the value of the adaptive feature interaction optimization loss function determined according to the th training process.

[0069] Step (5): Determine the historical gradient of the historical loss function value with respect to the historical weight matrix according to the historical loss function value and the historical weight matrix used by the encoder during the historical training process.

[0070] Step (6): Determine the conflict detection term of the target increment of the historical weight matrix according to the preset correction coefficient and the historical gradient; Among them, the conflict detection term is used to adjust the update direction of the weight matrix.

[0071] Specifically, the embodiments of the present application adopt a gradient conflict detection method. The conflict detection term detects the gradient change in the low-dimensional latent space during each iteration, ensuring that each gradient update moves in the correct direction. Especially when encountering gradient conflicts, it corrects the gradient direction, optimizes the learning rate selection, and avoids the model deviating from the optimal solution due to error accumulation.

[0072] For example, when abnormal events in the control system log of process data to be evaluated are input simultaneously with normal stress data, the gradient direction may conflict. The conflict detection term can adjust the update direction to avoid error accumulation and ensure stable optimization of the model in complex scenarios (such as mechanical records of multi-device collaborative operations); among them, the conflict detection term has the following formula: ; In the formula, represents the conflict detection term of the increment of the weight matrix determined according to the th training process; represents a preset correction coefficient; represents the gradient of the weight matrix of the encoder determined according to the th training process; represents the gradient of the weight matrix of the encoder determined according to the th training process.

[0073] Step (7): Determine the target increment according to the preset inertia parameter, preset learning rate, historical gradient, and conflict detection term; Among them, the inertia parameter is used to limit the influence of the update of the weight matrix in the historical training process on the target increment.

[0074] Step (8): Determine the target weight matrix used by the encoder and decoder in the current training process according to the historical weight matrix and the target increment.

[0075] Specifically, in actual operation, the traditional gradient descent method is prone to parameter update oscillations when there is noise data in the process data to be evaluated, and gradient conflicts (such as when abnormal events coexist with normal data) cause the model to deviate from the optimal solution.

[0076] To solve this problem, the embodiments of the present application solve this problem by providing an inertia calibration mechanism and a gradient conflict detection method; in the embodiments of the present application, different from traditional optimization methods (such as Adam, RMSProp) that only rely on gradient magnitude or exponential averaging, the parameter update of the autoencoder (including the parameters of the encoder and decoder) adopts an inertia calibration mechanism, and the historical gradient information is used to correct each update process, so that the autoencoder can update the parameters more stably and avoid the oscillation phenomenon in traditional gradient descent. For scenarios with a lot of noise such as mechanical operation records in the process data to be evaluated, for example, when there are irregular device start-stop signals in the control system log, the inertia calibration can smooth the parameter update path and avoid the model deviating from the optimal solution due to gradient mutation.

[0077] Among them, the target increment and the target weight matrix The calculation formula is as follows: ; ; In the formula, represents the increment of the weight matrix determined according to the th training process; represents a preset inertia coefficient; represents the learning rates of the preset encoder and decoder; represents the th loss function value of the autoencoder during the th training process with respect to the gradient of the weight matrix; represents the weight matrix used by the encoder during the th training process.

[0078] It should be noted that in the embodiments of the present application, the update method of the bias of the autoencoder can refer to the update method of the weight matrix, which will not be elaborated here.

[0079] S130: Input the compression data to be evaluated into the pre-trained heat treatment process performance evaluation model to obtain the evaluation result of the heat treatment process to be evaluated.

[0080] Specifically, in actual operation, the label should also be input into the heat treatment process performance evaluation model together with the training samples; in the embodiments of the present application, the content of the label may include different heat treatment qualification identifications, indicating whether the heat treatment process meets the expected process standard; in actual operation, the process standard can be divided into 5 levels, and the higher the level, the higher the process standard reached.

[0081] In the embodiments of the present application, the network structure of the heat treatment process performance evaluation model can be a deep neural network, a support vector machine, a random forest, a gradient boosting tree, or a decision tree, etc., and the present application does not make specific limitations on this.

[0082] In actual operation, when the network structure of the heat treatment process performance evaluation model is a deep neural network, the deep neural network may include: an input layer, a hidden layer, and an output layer.

[0083] Among them, the input layer is used to receive the process data to be evaluated.

[0084] The hidden layer includes: a fully connected layer, a batch normalization layer (BatchNorm), and a Dropout layer; the number of neurons in the fully connected layer is twice the dimension of the low-dimensional latent space, and the activation function of the fully connected layer is Leaky ReLU with a negative slope coefficient of 0.01 to alleviate the vanishing gradient; the batch normalization layer is used to accelerate convergence and suppress the shift of feature distribution; the dropout rate of the Dropout layer is 0.3 to prevent overfitting and improve the generalization ability to noisy data.

[0085] The output layer also includes a fully connected layer, which includes 5 neurons corresponding to the 5 process standards mentioned above, and its activation function is Softmax for outputting the probability distribution.

[0086] In actual operation, when the network structure of the heat treatment process performance evaluation model is a deep neural network, the loss function of the deep neural network can use the cross-entropy loss function; during its training process, the compressed feature set is divided into a training set, a validation set, and a test set according to 7:2:1; among them, the validation set is used for early stopping and hyperparameter tuning; during its training process, the deep neural network adopts the AdamW optimizer with an initial learning rate of 1e -3 , and a weight decay of 1e -4 , and dynamically adjusts the parameter update step size.

[0087] In actual operation, when the network structure of the heat treatment process performance evaluation model is a deep neural network, the deep neural network is used to output a 5-dimensional probability vector corresponding to each process data to be evaluated, representing the probability of belonging to each process standard. Further, the level corresponding to the maximum probability is taken as the final evaluation result; for example, if the 5-dimensional vector output by the deep neural network is [0.1, 0.05, 0.7, 0.1, 0.05], its final evaluation result should be the evaluation result corresponding to the probability value of 0.7.

[0088] In actual operation, in order to verify the superiority of the training method for the data compression model provided in the embodiments of the present application, that is, for the autoencoder, the present application now conducts the following comparative experiments between other existing training methods and the training method provided in the embodiments of the present application: As Figure 2 shown, Figure 2 is an example diagram of the training loss comparison of different training methods provided in the embodiments of the present application, Figure 2 the abscissa of which is the number of training iterations, Figure 2 and the ordinate is the reconstruction loss. This comparative experiment aims to verify the convergence stability of the inertial calibration gradient update strategy compared with traditional optimization methods. The experiment compares the changes in training loss of traditional gradient descent, momentum method, and the training method provided in the embodiments of the present application during 500 iterations.

[0089] The experimental results show that the loss curve of traditional gradient descent oscillates significantly. Although the momentum method has been improved, there are still periodic fluctuations. The training method provided in the embodiment of the present application incorporates historical gradient inertia calibration, and the loss curve decreases faster and smoother. The reconstruction loss is reduced under the same number of iterations. It can be seen that the inertia calibration mechanism provided in the embodiment of the present application effectively suppresses gradient oscillation by smoothing the parameter update path, thereby significantly improving the stability and convergence efficiency of model training.

[0090] like Figure 3 As shown, Figure 3 This is an example diagram of the effect of the autoencoder provided in the embodiment of the present application on the reconstruction accuracy. Figure 3 The horizontal axis of is the dimension of the low-dimensional latent space, Figure 3 The ordinate is the reconstruction error, which is represented by the mean square error (MSE). This experiment aims to explore the impact of the potential space dimension on the reconstruction accuracy and verify the adaptive advantages of the dynamic dimensionality expansion and contraction method.

[0091] The experimental results show that the experiment compares the reconstruction errors of a conventional autoencoder with a fixed dimension and the dynamic dimension method of the embodiment of the present application under different potential dimensions. The conventional method has a high error due to excessive information compression in low dimensions (<50), and overfitting due to redundant features in high dimensions (>80); while the embodiment of the present application maintains a stable low reconstruction error in the range of 10-100 dimensions by dynamically adjusting the dimension. It can be seen that the autoencoder provided by the embodiment of the present application can adaptively expand or compress the dimension according to the feature contribution, overcoming the defect that the dimension selection of the traditional method depends on prior knowledge, and significantly improving the adaptability of the model to complex data.

[0092] like Figure 4 As shown, Figure 4 This is an example diagram of the impact of gradient conflict detection on parameter updating provided by an embodiment of the present application. Figure 4 The horizontal axis is the number of iterations, Figure 4 The ordinate is the gradient direction consistency, which is represented by cosine similarity. This experiment verifies the optimization effect of the gradient conflict detection mechanism on parameter updating by analyzing the gradient direction consistency.

[0093] The experimental results show that the experiment compares the changes in cosine similarity of gradient updates with and without conflict detection, and finds that the gradient direction without conflict detection frequently deviates from the target (similarity <0.9), resulting in approximately 35% invalid parameter updates; while the embodiment of the present application dynamically corrects the gradient direction through the conflict detection item, so that the cosine similarity is stabilized above 0.9, thereby increasing the number of valid parameter updates. It can be seen that the conflict detection mechanism reduces random oscillations in the parameter update process by suppressing gradient direction conflicts, allowing the model to more reliably approach the optimal solution.

[0094] As shown Figure 5 in Figure 5 Fig. [figure number], this is a comparative example diagram of the ability of the energy flow to retain features in a noisy environment provided by the embodiment of the present application. Figure 5 The abscissa of [[figure]] is the noise level, Figure 5 and the ordinate of [[figure]] is the feature retention rate. This experiment evaluates the feature retention ability of the energy flow optimization method in a noisy environment, and compares the feature retention rate and signal-to-noise ratio of the traditional principal component analysis method and the embodiment of the present application at different noise levels.

[0095] The experimental results show that when the noise level exceeds 0.3, the principal component analysis method cannot distinguish noise from effective features, and the retention rate drops sharply to less than 70%; while the embodiment of the present application suppresses redundancy by dynamically quantifying the feature energy of the energy flow, and still maintains a feature retention rate of 82% under high noise (0.5), and the signal-to-noise ratio is increased by 6 - 8 dB. It can be seen that the energy flow optimization mechanism provided by the embodiment of the present application can effectively identify and enhance the representation of important features, and significantly improve the robustness of the model under noise interference.

[0096] Second, the present application provides a device for evaluating the heat treatment process performance of a safety hook based on a neural network. As shown Figure 6 in Figure 6 Fig. [figure number], this is a schematic structural diagram of the device for evaluating the heat treatment process performance of a safety hook based on a neural network provided by the embodiment of the present application. The device includes: a collection module 210, a compression module 220, and an evaluation module 230; The collection module 210 is configured to collect a variety of process data to be evaluated of the heat treatment process to be evaluated during the process of processing the safety hook by the heat treatment process to be evaluated; The compression module 220 is configured to input the variety of process data to be evaluated into a pre-trained data compression model to obtain the compressed data to be evaluated corresponding to the variety of process data to be evaluated; The evaluation module 230 is configured to input the compressed data to be evaluated into a pre-trained heat treatment process performance evaluation model to obtain the evaluation result of the heat treatment process to be evaluated.

[0097] In one implementation, the network structure of the data compression model is an autoencoder, and the autoencoder includes an encoder and a decoder; the device further includes: a training module; The training module is configured to perform iterative training on the autoencoder; Among them, during the process of performing iterative training on the autoencoder, the current training process before reaching the training stop condition includes: Determining the current sample input data corresponding to each training sample according to a plurality of training samples; Among them, the training samples include various sample process data, and the sample process data is collected during the sample heat treatment process for the safety hook; Input the current sample input data into the encoder to obtain the current low-dimensional latent space representation feature corresponding to the current sample input data; Input the current low-dimensional latent space representation feature into the decoder to obtain the current sample compressed data corresponding to the low-dimensional latent space representation feature.

[0098] In one implementation, the parameters of both the encoder and the decoder include weight matrices; before inputting the sample input data into the encoder to obtain the low-dimensional latent space representation feature corresponding to the sample input data, the training module is further configured to determine the historical loss function value of the autoencoder during the historical training process according to the historical latent space representation features output by the encoder during the historical training process; The training module is further configured to determine the historical gradient of the historical loss function value with respect to the historical weight matrix according to the historical loss function value and the historical weight matrix used by the encoder during the historical training process; The training module is further configured to determine the conflict detection term of the target increment of the historical weight matrix according to the preset correction coefficient and the historical gradient; Among them, the conflict detection term is used to adjust the update direction of the weight matrix; The training module is further configured to determine the target increment according to the preset inertia parameter, the preset learning rate, the historical gradient, and the conflict detection term; Among them, the inertia parameter is used to limit the influence of the update of the weight matrix during the historical training process on the target increment; The training module is further configured to determine the target weight matrices used by the encoder and the decoder during the current training process according to the historical weight matrix and the target increment.

[0099] In one implementation, the training module is further configured to determine the feature weight weighting coefficient, the energy flow loss function value, and the adaptive feature interaction optimization loss function value according to the historical latent space representation features; Among them, the feature weight weighting coefficient indicates the influence of the historical latent space representation feature on the target increment; the energy flow loss function value indicates the influence of the energy flow of the historical latent space representation feature on the target increment, and the energy flow indicates the information amount of the historical latent space representation feature; the adaptive feature interaction optimization loss function value indicates the influence of each dimension feature in the historical latent space representation feature on the target increment; The training module is further configured to determine the historical loss function value according to the historical sample input data input into the encoder during the historical training process, the preset sparsity weighting coefficient, the feature weight weighting coefficient, the energy flow loss function value, and the adaptive feature interaction optimization loss function value.

[0100] In one implementation, the training module is further configured to determine the feature contribution degree and information gain of the historical latent space representation features based on the historical latent space representation features; Among them, the feature contribution degree indicates the contribution of the historical latent space representation features to the historical sample compressed data output by the decoder during the historical training process; the information gain indicates the information quantity difference of the historical latent space representation features; The training module is further configured to determine the feature weight weighting coefficient according to a preset first adjustment coefficient, the historical latent space representation features, the feature contribution degree, and the information gain; Among them, the first adjustment coefficient is used to adjust the compression and reconstruction of the historical sample input data by the encoder and the decoder.

[0101] In one implementation, the training module is further configured to determine the variance and entropy of the low-dimensional latent space representation features based on the historical latent space representation features; among them, the entropy indicates the degree of dispersion of the low-dimensional latent space representation features; The training module is further configured to determine the feature contribution degree according to a preset second adjustment coefficient, the historical latent space representation features, the variance, and the entropy; Among them, the second adjustment coefficient is used to adjust the relative relationship between the variance and the entropy; The training module is further configured to determine the information gain according to the variance.

[0102] In one implementation, the training module is further configured to determine the energy flow and the deviation degree of the energy flow of the low-dimensional latent space representation features based on the historical latent space representation features; The training module is further configured to determine the energy flow loss function value according to a preset third adjustment coefficient, the energy flow, and the deviation degree; among them, the third adjustment coefficient is used to adjust the relative relationship between the energy flow and the entropy.

[0103] In one implementation, the training module is further configured to determine the feature interaction matrix according to the historical sample input data input to the encoder during the historical training process; Among them, the feature interaction matrix indicates the interaction result between multiple features included in the historical sample input data; The training module is further configured to perform normalization processing on the feature interaction matrix to obtain the normalized feature interaction matrix; The training module is further configured to determine the adaptive feature interaction optimization loss function value according to a preset weight coefficient, a regularization coefficient for adaptively optimizing the loss function value of the feature interaction 、the low-dimensional latent space representation features, and the historical weight matrix; Among them, the weight coefficient adjusts the influence of the low-dimensional latent space representation features on the autoencoder; the regularization coefficient is used to prevent overfitting.

[0104] Thirdly, the present application also provides an electronic device, including a memory, a processor, and a computer program stored on the memory and executable on the processor. When the processor executes the computer program, the steps of S110 to S130 provided in the above embodiments are implemented.

[0105] Fourthly, the present application also provides a computer-readable storage medium, on which a computer program is stored. When the computer program is run by a processor, the steps of S110 to S130 in the above embodiments are executed.

[0106] Fifthly, the computer program product provided by the present application includes a computer-readable storage medium storing program codes. The instructions included in the program codes can be used to execute the methods in the foregoing method embodiments. For the specific implementation, reference can be made to the steps of S110 to S130 in the method embodiments, which will not be elaborated herein.

[0107] In the embodiments provided by the present application, it should be understood that the disclosed device and method can be implemented in other ways. The device embodiments described above are merely illustrative. For example, the division of the units is only a logical function division, and there may be other division methods in actual implementation. For another example, multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the displayed or discussed mutual coupling or direct coupling or communication connection may be through some communication interfaces. The indirect coupling or communication connection of the device or unit can be in an electrical, mechanical or other form.

[0108] In addition, the units described as separate components may or may not be physically separated. The components displayed as units may or may not be physical units, that is, they may be located in one place, or may be distributed to multiple network units. Some or all of the units can be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0109] Furthermore, in each embodiment of the present application, the functional modules can be integrated together to form an independent part, or each module can exist alone, or two or more modules can be integrated to form an independent part.

[0110] It should be noted that if a function is implemented in the form of a software functional module and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present application, in essence, or the part that contributes to the prior art or a part of this technical solution can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions for causing a computer device (which may be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in various embodiments of the present application. The aforementioned storage medium includes: various media that can store program codes, such as USB flash drives, mobile hard disks, read-only memories (ROM), random access memories (RAM), magnetic disks, or optical discs.

[0111] In this document, relational terms such as first and second are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations.

[0112] The above are only the embodiments of the present application and are not used to limit the protection scope of the present application. For those skilled in the art, the present application can have various changes and modifications. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. A method for evaluating the performance of heat treatment process of safety hooks based on neural network, characterized in that: The method comprises: In the process of treating the safety hook with the heat treatment process to be evaluated, collecting a plurality of process data to be evaluated of the heat treatment process to be evaluated; Inputting the plurality of process data to be evaluated into a pre-trained data compression model to obtain compressed data to be evaluated corresponding to the plurality of process data to be evaluated; The compressed data to be evaluated is input into a pre-trained heat treatment process performance evaluation model to obtain an evaluation result of the heat treatment process to be evaluated.

2. The method according to claim 1, characterized in that The network structure of the data compression model is an autoencoder, and the autoencoder includes an encoder and a decoder; in the process of iterative training for the autoencoder, the current training process before reaching the training stop condition includes: According to a plurality of training samples, determining current sample input data corresponding to each training sample; The training samples include a variety of sample process data, and the sample process data is collected when the sample heat treatment process is performed on the safety hook; Inputting the current sample input data into the encoder to obtain a current low-dimensional latent space representation feature corresponding to the current sample input data; The current low-dimensional latent space representation feature is input into the decoder to obtain current sample compressed data corresponding to the low-dimensional latent space representation feature.

3. The method according to claim 2, characterized in that The parameters of the encoder and the decoder both include weight matrices; before inputting the current sample input data into the encoder to obtain a current low-dimensional latent space representation feature corresponding to the current sample input data, the method further includes: Determining a historical loss function value of the autoencoder in the historical training process according to historical latent space representation features output by the encoder in the historical training process; Determining a historical gradient of the historical loss function value relative to the historical weight matrix according to the historical loss function value and a historical weight matrix used by the encoder in a historical training process; Determining a conflict detection item of a target increment of the historical weight matrix according to a preset correction coefficient and the historical gradient; Wherein, the conflict detection item is used to adjust the update direction of the weight matrix; Determining the target increment according to a preset inertia parameter, a preset learning rate, the historical gradient and the conflict detection item; Wherein, the inertia parameter is used to limit the influence of the update of the weight matrix in the historical training process on the target increment; The target weight matrices used by the encoder and the decoder in the current training process are determined according to the historical weight matrix and the target increment.

4. The method according to claim 3, characterized in that: The determining, according to the historical latent space representation features output by the encoder in the historical training process, the historical loss function value of the autoencoder in the historical training process comprises: Determining feature weighting coefficients, energy flow loss function values, and adaptive feature interaction optimization loss function values ​​according to the historical latent space representation features; Among them, the feature weight coefficient indicates the influence of the historical latent space representation feature on the target increment; the energy flow loss function value indicates the influence of the energy flow of the historical latent space representation feature on the target increment, and the energy flow indicates the amount of information of the historical latent space representation feature; the adaptive feature interaction optimization loss function value indicates the influence of the features of each dimension in the historical latent space representation feature on the target increment; The historical loss function value is determined based on the historical sample input data input into the encoder during the historical training process, the preset sparsity weighting coefficient, the feature weighting coefficient, the energy flow loss function value and the adaptive feature interaction optimization loss function value.

5. The method according to claim 4, characterized in that Determining a feature weight coefficient according to the historical latent space representation feature includes: Determining, according to the historical latent space representation feature, a feature contribution and an information gain of the historical latent space representation feature; The feature contribution indicates the contribution of the historical latent space representation feature to the historical sample compression data output by the decoder during the historical training process; the information gain indicates the difference in the amount of information of the historical latent space representation feature; Determining the feature weighting coefficient according to a preset first adjustment coefficient, the historical latent space representation feature, the feature contribution and the information gain; The first adjustment coefficient is used to adjust the compression and reconstruction of the historical sample input data by the encoder and the decoder.

6. The method according to claim 5, characterized in that The step of determining the feature contribution and information gain of the historical latent space representation feature according to the historical latent space representation feature includes: Determining the variance and entropy of the low-dimensional latent space representation feature according to the historical latent space representation feature; wherein the entropy indicates the degree of discreteness of the low-dimensional latent space representation feature; Determining a feature contribution according to a preset second adjustment coefficient, the historical latent space representation feature, the variance, and the entropy; Wherein, the second adjustment coefficient is used to adjust the relative relationship between the variance and the entropy; The information gain is determined according to the variance.

7. The method according to claim 6, characterized in that Determining the energy flow loss function value according to the historical latent space representation feature includes: Determining an energy flow of the low-dimensional latent space representation feature and a deviation of the energy flow according to the historical latent space representation feature; The energy flow loss function value is determined according to a preset third adjustment coefficient, the energy flow and the deviation; wherein the third adjustment coefficient is used to adjust the relative relationship between the energy flow and the entropy.

8. The method according to claim 4, characterized in that The step of determining the adaptive feature interaction optimization loss function value according to the historical latent space representation feature includes: Determining a feature interaction matrix based on historical sample input data input to the encoder during historical training; Wherein, the feature interaction matrix indicates the interaction results between multiple features included in the historical sample input data; Performing normalization processing on the feature interaction matrix to obtain a normalized feature interaction matrix; Regularization coefficient of loss function value optimized interactively based on preset weight coefficient and preset adaptive feature , the low-dimensional latent space represents features and a historical weight matrix, and determines the adaptive feature interaction optimization loss function value; Among them, the weight coefficient The influence of the low-dimensional latent space representation feature on the autoencoder is adjusted; the regularization coefficient is used to prevent overfitting.

9. A device for evaluating the performance of heat treatment process of safety hooks based on neural network, characterized in that: The device comprises: an acquisition module, a compression module and an evaluation module; The acquisition module is used to collect a plurality of process data to be evaluated of the heat treatment process to be evaluated during the process of the safety hook being processed by the heat treatment process to be evaluated; The compression module is used to input the plurality of process data to be evaluated into a pre-trained data compression model to obtain compressed data to be evaluated corresponding to the plurality of process data to be evaluated; The evaluation module is used to input the compressed data to be evaluated into a pre-trained heat treatment process performance evaluation model to obtain an evaluation result of the heat treatment process to be evaluated.

10. The device according to claim 9, characterized in that The network structure of the data compression model is an autoencoder, and the autoencoder includes an encoder and a decoder; the device also includes: a training module; The training module is used to perform iterative training on the autoencoder; According to a plurality of training samples, determining current sample input data corresponding to each training sample; The training samples include a variety of sample process data, and the sample process data is collected when the sample heat treatment process is performed on the safety hook; Inputting the current sample input data into the encoder to obtain a current low-dimensional latent space representation feature corresponding to the current sample input data; The current low-dimensional latent space representation feature is input into the decoder to obtain current sample compressed data corresponding to the low-dimensional latent space representation feature.

Images