A new energy vehicle power battery health status detection method based on cell-to-battery pack migration

Abstract

The present invention relates to a method for detecting the health status of power batteries for new energy vehicles based on cell-to-battery pack migration, comprising: collecting voltage data of cell batteries during the cycle charging phase to construct a pre-trained cell battery SOH evaluation model; extracting the voltage curve of the battery pack during the charging phase; performing deep feature migration on the voltage curve of the battery pack during the charging phase to obtain target battery pack data; retraining the cell battery SOH evaluation model based on the target battery pack data to obtain the battery pack SOH evaluation model; and using the battery pack evaluation model to detect the health status of power batteries for new energy vehicles. The present invention designs an SOH evaluation method for migration from cell batteries to vehicle battery packs, optimizes the evaluation model using small-scale sample data, and achieves rapid evaluation of the battery pack SOH using a small-scale sample data set by selecting a suitable model main network and an innovative deep neuron feature optimization method.

Description

Technical Field

[0001] The present invention relates to the technical field of electric vehicle detection, and in particular to a method for detecting the health status of a power battery of a new energy vehicle based on cell-to-battery pack migration. Background Art

[0002] New energy vehicles (NEVs) have become a key development direction in the automotive industry due to their environmentally friendly and energy-efficient features. As a core component of NEVs, the battery pack's state of health (SOH) directly impacts the vehicle's range, charging speed, and safety. In actual use, the battery pack's SOH is affected by a variety of factors, including temperature, number of charge and discharge cycles, depth of discharge, and environmental conditions. These factors can cause battery capacity degradation and increase internal resistance, thereby impacting vehicle performance and posing safety risks. Accurate battery SOH assessment can not only predict battery performance degradation trends and provide maintenance recommendations to vehicle owners, but also improve the safety and reliability of the entire vehicle.

[0003] However, at present, there is a lack of direct means to detect the health status of vehicle battery packs, especially without unpacking them. The main reasons are as follows: First, lithium-ion batteries convert chemical energy into electrical energy through a series of electrochemical coupling reactions. Without unpacking, the battery pack is equivalent to a "black box" to the outside world, and it is difficult to detect the aging-related microchemical reactions inside it. In actual testing, indirect estimates can only be made through physical quantities such as voltage and current during the battery charging and discharging process, which brings challenges to detection; secondly, vehicle power-grade battery packs have the characteristics of high voltage, large capacity, and large volume. Therefore, data collection for automotive-grade battery packs is difficult and costly. The new development situation requires a faster, more accurate, and more intelligent vehicle battery pack health status assessment method.

[0004] The continuous advancement of big data science and the widespread application of deep learning technology in the engineering field have provided an advanced operational data-driven approach for vehicle battery pack SOH assessment, meeting new requirements for assessment methods.

[0005] The latest methods are based on acquiring characteristic data during the battery charging process, such as current, voltage, and temperature, and evaluating the battery's health status through analysis and modeling. This type of method is called a data-driven approach. The general process is divided into: 1. Data acquisition. Current, voltage, temperature, and other data are collected from the battery system during battery charging; 2. Data preprocessing. The collected data is cleaned, corrected, and preprocessed to ensure data consistency and accuracy; 3. Feature extraction. Features related to the battery's health status are extracted from the processed data based on deep learning or machine learning; 4. Modeling. The extracted features are used to build a model, usually a mathematical model that associates the battery's state of health (SOH) with the features. 5. Calculation results. The battery's state of health is calculated using the model. Figure 1 This is a flowchart of the data-driven method.

[0006] The above method has been successfully applied to the health status testing of small single-cell batteries, such as 18650 batteries, and has achieved ideal evaluation results. However, when the evaluation object is changed to a larger-capacity automotive-grade battery pack, the difference in data distribution between the single cell and the battery pack leads to reduced evaluation accuracy or model failure. In particular, due to the wide variety of battery packs on the market and their complex chemical structures, the model must be retrained by collecting operating data from the battery pack to be tested in order to fully utilize the performance of the evaluation model. In addition, due to the high voltage and large capacity of the battery pack, obtaining the operating data required for retraining is time-consuming and labor-intensive, which seriously hinders the successful application of this method.

[0007] How to improve the model's cross-object evaluation capabilities and shorten the time and effort required for model retraining are key issues in ultimately realizing the practical application of data-driven methods in the health status assessment of battery packs in vehicles. Summary of the Invention

[0008] The purpose of the present invention is to propose a new energy vehicle power battery health status detection method based on cell-to-battery pack migration to solve the problems existing in the above-mentioned prior art. The present invention designs an SOH evaluation method for migration from battery cells to vehicle battery packs, and successfully extends the method that could only be applied to battery cell SOH detection in the past to the field of vehicle detection. The evaluation model is optimized using small-scale sample data, and by selecting a suitable model main network and an innovative deep neuron feature optimization method, finally, based on the battery cell SOH evaluation model, a rapid evaluation of the battery pack SOH is achieved through a small-scale sample data set.

[0009] To achieve the above object, the present invention provides the following solutions:

[0010] A method for detecting the health status of a power battery of a new energy vehicle based on cell-to-battery pack migration includes:

[0011] Collecting voltage data of a single cell during the cycle charge phase to construct a pre-trained single cell SOH assessment model; wherein the chemical composition of the single cell is the same as that of the battery pack;

[0012] Extract the voltage curve of the battery pack during the charging phase;

[0013] Performing deep feature migration on the voltage curve of the battery pack during the charging phase to obtain target battery pack data;

[0014] Retraining the single cell SOH evaluation model based on the target battery pack data to obtain a battery pack SOH evaluation model;

[0015] The battery pack SOH evaluation model is used to detect the health status of power batteries in new energy vehicles.

[0016] Optionally, collecting voltage data of a single battery during a cycle charging phase includes:

[0017] Prepare battery cells with the same chemical composition as the battery pack as experimental batteries;

[0018] Performing a charge and discharge experiment on the experimental battery;

[0019] The maximum dischargeable capacity of the battery after charge and discharge is detected. If the ratio of the maximum discharge capacity to the rated capacity is higher than a preset threshold, the charge and discharge experiment is repeated; when the ratio of the maximum discharge capacity to the rated capacity is lower than the preset threshold, all voltage data curves of the charging stage are extracted.

[0020] Optionally, performing a charge and discharge experiment on the experimental battery includes:

[0021] The battery cell is charged at a preset rate constant current until the maximum allowed charging voltage is reached, and then discharged at the preset rate constant current until the maximum allowed discharging voltage is reached.

[0022] Optionally, establishing a pre-trained single-cell battery SOH evaluation model includes:

[0023] Build an initial network model based on the long short-term memory neural network model and the three-layer fully connected neural network model;

[0024] Standardize the voltage data of single battery during cycle charging stage;

[0025] The standardized voltage data is input into the initial network model, and the initial network model is trained to obtain a single cell SOH evaluation model.

[0026] Optionally, the expression for normalizing the voltage data of a single battery during the cycle charging stage is:

[0027] x' = (x - μ) / σ

[0028] Where x' is the normalized voltage data, x is the original data, μ is the mean of the feature, and σ is the standard deviation of the feature.

[0029] Optionally, inputting the standardized voltage data into the initial network model includes:

[0030] Use long short-term memory neural network to extract temporal features from data;

[0031] A three-layer fully connected neural network is used to extract deep neuron features in the data.

[0032] Optionally, extracting temporal features from data using a long short-term memory neural network includes:

[0033] The standardized voltage data is organized into a three-dimensional tensor, which is the number of samples, time steps, and number of features. The organized data is cut in the form of a sliding window;

[0034] The cut data are put into the long short-term memory network in chronological order. The hidden state of the current time step is calculated based on the hidden state of the previous time step and the input of the current time step, thereby extracting the time series features.

[0035] Optionally, a method for performing deep feature migration on the voltage curve in the charging stage is:

[0036] U'=U×U 单max / U 包max

[0037] Among them, U' is the characteristic after migration, U is the characteristic of the battery pack before migration, and U 单max is the maximum charging voltage of the monomer, U 包max The maximum charging voltage of the battery pack.

[0038] Optionally, retraining the single cell SOH evaluation model includes:

[0039] Preprocessing the single cell SOH evaluation model;

[0040] Retraining the single cell SOH evaluation model using the preprocessed model;

[0041] Among them, the preprocessing includes: freezing the neurons of the long short-term memory network model and only keeping the neurons of the full neural network model activated.

[0042] The beneficial effects of the present invention are:

[0043] Compared with the existing methods, the advantages of the present invention are mainly reflected in the following three aspects: First, the present invention constructs an SOH pre-training model based on the cycle aging data of small battery cells, and then migrates the battery pack data to make it have the same distribution as the cell data, so that the model can effectively detect the SOH of the battery pack. Compared with the method of directly using battery pack data to build a model, the labor and time costs of collecting data are effectively reduced. Second, by retraining the pre-trained model with a small amount of battery pack data, the accuracy of the model's evaluation of the battery pack's SOH can be effectively improved, and a battery pack SOH evaluation model can be established quickly and at low cost. Third, a battery SOH detection model is constructed based on a large artificial intelligence model to extract the conversion relationship between battery data and battery SOH, which achieves a higher accuracy rate than traditional models. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0045] Figure 1 This is a schematic diagram of the data-driven approach;

[0046] Figure 2 This is a flow chart of a method for detecting the health status of a power battery of a new energy vehicle based on migration from a single cell to a battery pack according to an embodiment of the present invention;

[0047] Figure 3 The figure is a flow chart of a method for detecting the health status of a new energy vehicle power battery applied to an EV power battery testing station according to an embodiment of the present invention. DETAILED DESCRIPTION

[0048] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0049] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0050] like Figure 2 As shown, this embodiment proposes a method for detecting the health status of a power battery of a new energy vehicle based on migration from a single cell to a battery pack, including:

[0051] Collect voltage data from single-cell batteries during the cycle charge phase to build a pre-trained single-cell SOH assessment model; the chemical composition of the single-cell battery is the same as that of the battery pack;

[0052] Extract the voltage curve of the battery pack during the charging phase;

[0053] Perform deep feature migration on the voltage curve of the battery pack during the charging phase to obtain target battery pack data;

[0054] Based on the target battery pack data, the single cell SOH evaluation model is retrained to obtain the battery pack SOH evaluation model;

[0055] The battery pack SOH assessment model is used to detect the health status of power batteries in new energy vehicles.

[0056] Specifically, this embodiment first establishes a pre-trained health assessment model based on single cells. Then, using a deep feature transfer method, the pre-trained model is retrained using a small amount of battery pack data with battery pack features. Ultimately, a transfer learning model is obtained that can effectively assess the battery pack's state of health (SOH). Specifically, in the first stage, a single cell with the same chemical composition as the battery pack to be tested is prepared as a test cell. Due to the small capacity and low voltage of the single cell, a rapid aging test can be performed on it with minimal labor and time costs to collect aging data. The cell is charged at a constant current rate of C1 to the maximum allowable charge voltage, then discharged at a constant current rate of C1 to the maximum allowable discharge voltage. The maximum dischargeable capacity of the battery is measured at this point. If the ratio of maximum discharge capacity to rated capacity (SOH) is greater than 80%, the charge and discharge test is repeated. When the SOH is less than 80%, all voltage curves during the charging phase are extracted as a training database. A pre-trained single cell SOH assessment model is established based on the training database. In the second stage, communication with the vehicle is performed to obtain the rated capacity Qr and maximum charging voltage Umax of the battery pack. According to I=Qr*C1, the charging current I is calculated. Then, the battery pack is charged with constant current I to the maximum charging voltage, and the voltage curve of the charging stage is extracted. The extracted battery pack voltage curve is deeply characterized and distributed with the battery cell data. The specific method is U'=U×U 单 max / U 包 max, where U' is the feature after migration, U is the feature of the battery pack before migration, and U 单 max is the maximum charging voltage of the monomer, U 包max is the maximum charging voltage of the battery pack. After migrating the data, the model is retrained based on a small amount of battery pack data. The retraining method freezes some neurons used for time series feature extraction in the model, keeping deeper neurons activated. After importing a small amount of battery pack features, the battery pack SOH assessment model is obtained.

[0057] Furthermore, collecting voltage data of a single battery during the cycle charging phase includes:

[0058] Prepare battery cells with the same chemical composition as the battery pack as experimental batteries;

[0059] Carry out charge and discharge experiments on the experimental battery;

[0060] The maximum discharge capacity of the battery after charge and discharge is detected. If the ratio of the maximum discharge capacity to the rated capacity is higher than the preset threshold, the charge and discharge experiment is repeated; when the ratio of the maximum discharge capacity to the rated capacity is lower than the preset threshold, all voltage data curves in the charging stage are extracted.

[0061] Furthermore, the charge and discharge experiments on the experimental battery include:

[0062] The battery cells are charged at a preset constant current rate until the maximum allowable charging voltage is reached, and then discharged at a preset constant current rate until the maximum allowable discharge voltage is reached.

[0063] Specifically, this embodiment is applied to EV power battery testing station; its technical process is as follows Figure 3 As shown in the figure, during the data acquisition phase, battery cells with the same chemical composition as the vehicle under test are prepared and subjected to constant current cyclic charge and discharge at a current rate of 0.5C until their maximum available capacity falls below 80% of the rated available capacity. Voltage data from all charging phases is collected to form the training dataset. An EV battery pack data reader is used to read the battery pack information of the vehicle under test, obtaining its rated capacity and maximum charging voltage. The vehicle under test is also charged and discharged once at a current rate of 0.5C, and voltage data from the charging phase is collected to form the test data.

[0064] Furthermore, establishing a pre-trained single-cell battery SOH evaluation model includes:

[0065] Build an initial network model based on the long short-term memory neural network model and the three-layer fully connected neural network model;

[0066] Standardize the voltage data of single battery during cycle charging stage;

[0067] The standardized voltage data is input into the initial network model, the initial network model is trained, and a single cell SOH evaluation model is obtained.

[0068] Furthermore, inputting the normalized voltage data into the initial network model includes:

[0069] Use long short-term memory neural network to extract temporal features from data;

[0070] A three-layer fully connected neural network is used to extract deep neuron features in the data.

[0071] Furthermore, the long short-term memory neural network is used to extract temporal features from the data, including:

[0072] The standardized voltage data is organized into a three-dimensional tensor, which is the number of samples, time steps, and number of features. The organized data is cut in the form of a sliding window;

[0073] The cut data are put into the long short-term memory network in chronological order. The hidden state of the current time step is calculated based on the hidden state of the previous time step and the input of the current time step, thereby extracting the time series features.

[0074] Furthermore, retraining the single cell SOH evaluation model includes:

[0075] Preprocessing of single cell battery SOH evaluation model;

[0076] Use the preprocessed model to retrain the single cell SOH evaluation model;

[0077] Among them, the preprocessing includes: freezing the neurons of the long short-term memory network model and only keeping the neurons of the full neural network model activated.

[0078] Specifically, in this embodiment, feature migration: based on the formula U'=U×U 单max / U 包max , perform feature migration on the battery pack data to make it have the same distribution as the cell data. Where U' is the feature after migration, U is the feature of the battery pack before migration, and U 单max is the maximum charging voltage of the monomer, U 包max The maximum charging voltage of the battery pack.

[0079] Specifically, in this embodiment, standardization: the extracted features are standardized based on the formula x'=(x-μ) / σ, where x is the original data, μ is the mean of the feature, and σ is the standard deviation of the feature.

[0080] Specifically, in this embodiment, temporal feature extraction is performed on the standardized features. First, the raw data is organized into a three-dimensional tensor, representing the number of samples, the time step (=600), and the number of features (=1). The entire data segment is then segmented using a sliding window with a window size of 10 and a step size of 10. The segmented data is then fed into a long short-term memory network in chronological order. The hidden state of the current time step is calculated from the hidden state of the previous time step and the input of the current time step, thereby extracting temporal features.

[0081] Specifically, in this embodiment, deep neuron feature extraction and migration: After the long short-term memory network, a three-level full neural network is constructed to extract deep neuron features. The number of neurons in each layer of the network is 100, 100, and 50 respectively. When building a pre-trained model with battery cell data, all neurons in the long short-term memory network and the full neural network are kept activated. When retraining the model with battery pack data, the long short-term memory network neurons are frozen, and only the full neural network neurons are kept activated.

[0082] Specifically, in this embodiment, a SOH evaluation model is constructed: all data are input into the constructed model, and the SOH evaluation result is output.

[0083] This embodiment is applied in an EV battery pack inspection station. Through this method, the SOH of the vehicle battery pack can be quickly and effectively evaluated.

[0084] The following uses the application in the annual vehicle inspection station as an example to illustrate the detailed implementation process of the technical solution of this embodiment.

[0085] Fixed vehicle testing equipment equipped with a high-voltage charging gun, current sensor, voltage sensor, memory, and processor.

[0086] Data Collection: At the annual vehicle inspection station, the battery health monitoring equipment is connected to the battery system of the vehicle being inspected. A high-voltage charging gun is used to charge the vehicle at a high rate (80A) to ensure that the battery quickly reaches a stable charging state. During the charging phase, battery pack data, including current, voltage, and temperature, is collected and saved as raw data for the model input.

[0087] Model Construction: An electric vehicle battery health status detection model is built using a long short-term memory neural network (LSTM) and a three-layer fully connected neural network architecture. Raw data is normalized to ensure it is suitable for model input. The preprocessed data is then fed into the network. The LSTM neural network extracts temporal features from the data, while the fully connected neural network extracts deep neural network features from the data, ultimately outputting the SOH.

[0088] Parameter determination: Using pre-prepared battery cells of the same chemical type, the charge rate is calculated by dividing 80A by the rated capacity of the battery pack to be tested. The battery is then charged at a constant current at this charge rate to obtain charging voltage data. This data is input into the model to obtain a predicted SOH value. The mean square error between the predicted SOH value and the actual SOH value of the battery cell is used as the model training loss function. The model parameters are adjusted using the backpropagation algorithm to minimize the difference between the model output and the actual SOH. After the parameters are determined, the model is deployed in the device memory before the test equipment leaves the factory.

[0089] Battery State of Health (SOH) testing: Using battery health testing equipment, the same C1 rate constant current is applied to the vehicle under test to obtain the battery pack data at that time. Based on the formula U' = U × U single max / U pack max, the battery pack data is feature-shifted to align with the cell data. Here, U' represents the post-migration features, U represents the pre-migration battery pack features, U single max represents the maximum cell charge voltage, and U pack max represents the maximum battery pack charge voltage. The migrated features are input into a parameterized model to obtain an estimate of the vehicle's state of health (SOH).

[0090] Application: This technology enables rapid and accurate battery health assessment at annual vehicle inspection stations, ensuring that electric vehicle battery systems meet annual inspection standards. This helps improve inspection station efficiency and ensures road safety. It also provides a convenient way for vehicle owners to monitor their vehicle's battery health status and take necessary maintenance measures.

[0091] Compared with the existing methods, the advantages of this embodiment are mainly reflected in the following three aspects: First, this method constructs an SOH pre-training model based on the cycle aging data of small battery cells, and then migrates the battery pack data to make it have the same distribution as the cell data, so that the model can effectively detect the SOH of the battery pack. Compared with the method of directly using battery pack data to build a model, the labor and time costs of collecting data are effectively reduced. Second, by retraining the pre-trained model with a small amount of battery pack data, the accuracy of the model's evaluation of the battery pack's SOH can be effectively improved, and a battery pack SOH evaluation model can be established quickly and at low cost. Third, a battery SOH detection model is constructed based on a large artificial intelligence model to extract the conversion relationship between battery data and battery SOH, which achieves a higher accuracy rate than traditional models.

[0092] This embodiment proposes a novel vehicle battery pack SOH detection method, which constructs an SOH evaluation model through battery cells and migrates battery cell to battery pack data, effectively detecting the SOH of the battery pack inside the vehicle without unpacking.

[0093] By retraining and optimizing the model with a small amount of battery pack sample data, the evaluation performance of the SOH assessment model for battery pack data is effectively improved, expanding the application of data-driven methods in the field of vehicle inspection.

[0094] The embodiments described above are merely descriptions of preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Without departing from the spirit of the present invention, various modifications and improvements made to the technical solutions of the present invention by persons skilled in the art should fall within the scope of protection defined by the claims of the present invention.

Claims

1. A method for detecting the health status of a new energy vehicle power battery based on cell-to-battery pack migration, characterized in that: include: Collecting voltage data of a single cell during the cycle charge phase to construct a pre-trained single cell SOH assessment model; wherein the chemical composition of the single cell is the same as that of the battery pack; Extract the voltage curve of the battery pack during the charging phase; Performing deep feature migration on the voltage curve of the battery pack during the charging phase to obtain target battery pack data; The method for performing deep feature migration on the voltage curve in the charging stage is as follows: U'=U×U 单max / IN 包max Among them, U' is the characteristic after migration, U is the characteristic of the battery pack before migration, and U 单max is the maximum charging voltage of the monomer, U 包max is the maximum charging voltage of the battery pack; Retraining the single cell SOH evaluation model based on the target battery pack data to obtain a battery pack SOH evaluation model; The battery pack SOH evaluation model is used to detect the health status of power batteries in new energy vehicles.

2. The method for detecting the health status of a power battery of a new energy vehicle based on cell-to-battery pack migration according to claim 1, characterized in that: The voltage data collected during the cycle charging phase of a single battery includes: Prepare battery cells with the same chemical composition as the battery pack as experimental batteries; Performing a charge and discharge experiment on the experimental battery; The maximum dischargeable capacity of the battery after charge and discharge is detected. If the ratio of the maximum discharge capacity to the rated capacity is higher than a preset threshold, the charge and discharge experiment is repeated; when the ratio of the maximum discharge capacity to the rated capacity is lower than the preset threshold, all voltage data curves of the charging stage are extracted.

3. The method for detecting the health status of a power battery of a new energy vehicle based on cell-to-battery pack migration according to claim 2, characterized in that: The charge and discharge experiment of the experimental battery includes: The battery cell is charged at a preset rate constant current until the maximum allowed charging voltage is reached, and then discharged at the preset rate constant current until the maximum allowed discharging voltage is reached.

4. The method for detecting the health status of a power battery of a new energy vehicle based on cell-to-battery pack migration according to claim 1, characterized in that: Establishing a pre-trained single-cell battery SOH evaluation model includes: Construct an initial network model based on the long short-term memory neural network model and the three-layer fully connected neural network model; Standardize the voltage data of single battery during the cycle charging stage; The standardized voltage data is input into the initial network model, and the initial network model is trained to obtain a single cell SOH evaluation model.

5. The method for detecting the health status of a power battery of a new energy vehicle based on cell-to-battery pack migration according to claim 4, characterized in that: The expression for normalizing the voltage data of a single battery during the cycle charging phase is: x' = (x - μ) / σ Where x' is the normalized voltage data, x is the original data, μ is the mean of the feature, and σ is the standard deviation of the feature.

6. The method for detecting the health status of a power battery of a new energy vehicle based on cell-to-battery pack migration according to claim 4, characterized in that: Inputting the normalized voltage data into the initial network model comprises: Use long short-term memory neural network to extract temporal features from data; A three-layer fully connected neural network is used to extract deep neuron features in the data.

7. The method for detecting the health status of a power battery of a new energy vehicle based on cell-to-battery pack migration according to claim 6, characterized in that: The time series features extracted from data using long short-term memory neural networks include: The standardized voltage data is organized into a three-dimensional tensor, which is the number of samples, time steps, and number of features. The organized data is cut in the form of a sliding window; The cut data are put into the long short-term memory network in chronological order. The hidden state of the current time step is calculated based on the hidden state of the previous time step and the input of the current time step, thereby extracting the time series features.

8. The method for detecting the health status of a power battery of a new energy vehicle based on cell-to-battery pack migration according to claim 4, characterized in that: Retraining the single cell SOH evaluation model includes: Preprocessing the single cell SOH evaluation model; Retraining the single cell SOH evaluation model using the preprocessed model; Among them, the preprocessing includes: freezing the neurons of the long short-term memory network model and only keeping the neuron activations of the full neural network model.

Images